freebsd-dev/sys/dev/txp/if_txp.c

2990 lines
86 KiB
C
Raw Normal View History

/* $OpenBSD: if_txp.c,v 1.48 2001/06/27 06:34:50 kjc Exp $ */
/*-
* Copyright (c) 2001
* Jason L. Wright <jason@thought.net>, Theo de Raadt, and
* Aaron Campbell <aaron@monkey.org>. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* 3. All advertising materials mentioning features or use of this software
* must display the following acknowledgement:
* This product includes software developed by Jason L. Wright,
* Theo de Raadt and Aaron Campbell.
* 4. Neither the name of the author nor the names of any co-contributors
* may be used to endorse or promote products derived from this software
* without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHORS ``AS IS'' AND ANY EXPRESS OR
* IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
* WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL Bill Paul OR THE VOICES IN HIS HEAD
* BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
* CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
* SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
* INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
* CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
* ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
* THE POSSIBILITY OF SUCH DAMAGE.
*/
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
/*
* Driver for 3c990 (Typhoon) Ethernet ASIC
*/
#include <sys/param.h>
#include <sys/systm.h>
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
#include <sys/bus.h>
#include <sys/endian.h>
#include <sys/kernel.h>
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
#include <sys/lock.h>
#include <sys/malloc.h>
#include <sys/mbuf.h>
2004-05-30 20:08:47 +00:00
#include <sys/module.h>
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
#include <sys/mutex.h>
#include <sys/queue.h>
#include <sys/rman.h>
#include <sys/socket.h>
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
#include <sys/sockio.h>
#include <sys/sysctl.h>
#include <sys/taskqueue.h>
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
#include <net/bpf.h>
#include <net/if.h>
#include <net/if_var.h>
#include <net/if_arp.h>
#include <net/ethernet.h>
#include <net/if_dl.h>
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
#include <net/if_media.h>
#include <net/if_types.h>
#include <net/if_vlan_var.h>
#include <netinet/in.h>
#include <netinet/in_systm.h>
#include <netinet/ip.h>
#include <dev/mii/mii.h>
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
#include <dev/pci/pcireg.h>
#include <dev/pci/pcivar.h>
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
#include <machine/bus.h>
#include <machine/in_cksum.h>
#include <dev/txp/if_txpreg.h>
#include <dev/txp/3c990img.h>
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
MODULE_DEPEND(txp, pci, 1, 1, 1);
MODULE_DEPEND(txp, ether, 1, 1, 1);
/*
* XXX Known Typhoon firmware issues.
*
* 1. It seems that firmware has Tx TCP/UDP checksum offloading bug.
* The firmware hangs when it's told to compute TCP/UDP checksum.
* I'm not sure whether the firmware requires special alignment to
* do checksum offloading but datasheet says nothing about that.
* 2. Datasheet says nothing for maximum number of fragmented
* descriptors supported. Experimentation shows up to 16 fragment
* descriptors are supported in the firmware. For TSO case, upper
* stack can send 64KB sized IP datagram plus link header size(
* ethernet header + VLAN tag) frame but controller can handle up
* to 64KB frame given that PAGE_SIZE is 4KB(i.e. 16 * PAGE_SIZE).
* Because frames that need TSO operation of hardware can be
* larger than 64KB I disabled TSO capability. TSO operation for
* less than or equal to 16 fragment descriptors works without
* problems, though.
* 3. VLAN hardware tag stripping is always enabled in the firmware
* even if it's explicitly told to not strip the tag. It's
* possible to add the tag back in Rx handler if VLAN hardware
* tag is not active but I didn't try that as it would be
* layering violation.
* 4. TXP_CMD_RECV_BUFFER_CONTROL does not work as expected in
* datasheet such that driver should handle the alignment
* restriction by copying received frame to align the frame on
* 32bit boundary on strict-alignment architectures. This adds a
* lot of CPU burden and it effectively reduce Rx performance on
* strict-alignment architectures(e.g. sparc64, arm, mips and ia64).
*
* Unfortunately it seems that 3Com have no longer interests in
* releasing fixed firmware so we may have to live with these bugs.
*/
#define TXP_CSUM_FEATURES (CSUM_IP)
/*
* Various supported device vendors/types and their names.
*/
static struct txp_type txp_devs[] = {
{ TXP_VENDORID_3COM, TXP_DEVICEID_3CR990_TX_95,
"3Com 3cR990-TX-95 Etherlink with 3XP Processor" },
{ TXP_VENDORID_3COM, TXP_DEVICEID_3CR990_TX_97,
"3Com 3cR990-TX-97 Etherlink with 3XP Processor" },
{ TXP_VENDORID_3COM, TXP_DEVICEID_3CR990B_TXM,
"3Com 3cR990B-TXM Etherlink with 3XP Processor" },
{ TXP_VENDORID_3COM, TXP_DEVICEID_3CR990_SRV_95,
"3Com 3cR990-SRV-95 Etherlink Server with 3XP Processor" },
{ TXP_VENDORID_3COM, TXP_DEVICEID_3CR990_SRV_97,
"3Com 3cR990-SRV-97 Etherlink Server with 3XP Processor" },
{ TXP_VENDORID_3COM, TXP_DEVICEID_3CR990B_SRV,
"3Com 3cR990B-SRV Etherlink Server with 3XP Processor" },
{ 0, 0, NULL }
};
static int txp_probe(device_t);
static int txp_attach(device_t);
static int txp_detach(device_t);
static int txp_shutdown(device_t);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
static int txp_suspend(device_t);
static int txp_resume(device_t);
static int txp_intr(void *);
static void txp_int_task(void *, int);
static void txp_tick(void *);
static int txp_ioctl(struct ifnet *, u_long, caddr_t);
static void txp_start(struct ifnet *);
static void txp_start_locked(struct ifnet *);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
static int txp_encap(struct txp_softc *, struct txp_tx_ring *, struct mbuf **);
static void txp_stop(struct txp_softc *);
static void txp_init(void *);
static void txp_init_locked(struct txp_softc *);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
static void txp_watchdog(struct txp_softc *);
2002-03-20 02:08:01 +00:00
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
static int txp_reset(struct txp_softc *);
static int txp_boot(struct txp_softc *, uint32_t);
static int txp_sleep(struct txp_softc *, int);
static int txp_wait(struct txp_softc *, uint32_t);
2002-03-20 02:08:01 +00:00
static int txp_download_fw(struct txp_softc *);
static int txp_download_fw_wait(struct txp_softc *);
static int txp_download_fw_section(struct txp_softc *,
2002-03-20 02:08:01 +00:00
struct txp_fw_section_header *, int);
static int txp_alloc_rings(struct txp_softc *);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
static void txp_init_rings(struct txp_softc *);
static int txp_dma_alloc(struct txp_softc *, char *, bus_dma_tag_t *,
bus_size_t, bus_size_t, bus_dmamap_t *, void **, bus_size_t, bus_addr_t *);
static void txp_dma_free(struct txp_softc *, bus_dma_tag_t *, bus_dmamap_t *,
void **);
static void txp_free_rings(struct txp_softc *);
2002-03-20 02:08:01 +00:00
static int txp_rxring_fill(struct txp_softc *);
static void txp_rxring_empty(struct txp_softc *);
static void txp_set_filter(struct txp_softc *);
static int txp_cmd_desc_numfree(struct txp_softc *);
static int txp_command(struct txp_softc *, uint16_t, uint16_t, uint32_t,
uint32_t, uint16_t *, uint32_t *, uint32_t *, int);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
static int txp_ext_command(struct txp_softc *, uint16_t, uint16_t,
uint32_t, uint32_t, struct txp_ext_desc *, uint8_t,
2002-03-20 02:08:01 +00:00
struct txp_rsp_desc **, int);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
static int txp_response(struct txp_softc *, uint16_t, uint16_t,
2002-03-20 02:08:01 +00:00
struct txp_rsp_desc **);
static void txp_rsp_fixup(struct txp_softc *, struct txp_rsp_desc *,
2002-03-20 02:08:01 +00:00
struct txp_rsp_desc *);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
static int txp_set_capabilities(struct txp_softc *);
2002-03-20 02:08:01 +00:00
static void txp_ifmedia_sts(struct ifnet *, struct ifmediareq *);
static int txp_ifmedia_upd(struct ifnet *);
#ifdef TXP_DEBUG
2002-03-20 02:08:01 +00:00
static void txp_show_descriptor(void *);
#endif
2002-03-20 02:08:01 +00:00
static void txp_tx_reclaim(struct txp_softc *, struct txp_tx_ring *);
static void txp_rxbuf_reclaim(struct txp_softc *);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
#ifndef __NO_STRICT_ALIGNMENT
static __inline void txp_fixup_rx(struct mbuf *);
#endif
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
static int txp_rx_reclaim(struct txp_softc *, struct txp_rx_ring *, int);
static void txp_stats_save(struct txp_softc *);
static void txp_stats_update(struct txp_softc *, struct txp_rsp_desc *);
static void txp_sysctl_node(struct txp_softc *);
static int sysctl_int_range(SYSCTL_HANDLER_ARGS, int, int);
static int sysctl_hw_txp_proc_limit(SYSCTL_HANDLER_ARGS);
static int prefer_iomap = 0;
TUNABLE_INT("hw.txp.prefer_iomap", &prefer_iomap);
static device_method_t txp_methods[] = {
/* Device interface */
DEVMETHOD(device_probe, txp_probe),
DEVMETHOD(device_attach, txp_attach),
DEVMETHOD(device_detach, txp_detach),
DEVMETHOD(device_shutdown, txp_shutdown),
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
DEVMETHOD(device_suspend, txp_suspend),
DEVMETHOD(device_resume, txp_resume),
{ NULL, NULL }
};
static driver_t txp_driver = {
"txp",
txp_methods,
sizeof(struct txp_softc)
};
static devclass_t txp_devclass;
DRIVER_MODULE(txp, pci, txp_driver, txp_devclass, 0, 0);
static int
2009-03-11 08:25:18 +00:00
txp_probe(device_t dev)
{
struct txp_type *t;
t = txp_devs;
2009-03-11 08:49:17 +00:00
while (t->txp_name != NULL) {
if ((pci_get_vendor(dev) == t->txp_vid) &&
(pci_get_device(dev) == t->txp_did)) {
device_set_desc(dev, t->txp_name);
2009-03-11 08:49:17 +00:00
return (BUS_PROBE_DEFAULT);
}
t++;
}
2009-03-11 08:49:17 +00:00
return (ENXIO);
}
static int
2009-03-11 08:25:18 +00:00
txp_attach(device_t dev)
{
struct txp_softc *sc;
struct ifnet *ifp;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
struct txp_rsp_desc *rsp;
uint16_t p1;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
uint32_t p2, reg;
int error = 0, pmc, rid;
uint8_t eaddr[ETHER_ADDR_LEN], *ver;
sc = device_get_softc(dev);
sc->sc_dev = dev;
mtx_init(&sc->sc_mtx, device_get_nameunit(dev), MTX_NETWORK_LOCK,
MTX_DEF);
callout_init_mtx(&sc->sc_tick, &sc->sc_mtx, 0);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
TASK_INIT(&sc->sc_int_task, 0, txp_int_task, sc);
TAILQ_INIT(&sc->sc_busy_list);
TAILQ_INIT(&sc->sc_free_list);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
ifmedia_init(&sc->sc_ifmedia, 0, txp_ifmedia_upd, txp_ifmedia_sts);
ifmedia_add(&sc->sc_ifmedia, IFM_ETHER | IFM_10_T, 0, NULL);
ifmedia_add(&sc->sc_ifmedia, IFM_ETHER | IFM_10_T | IFM_HDX, 0, NULL);
ifmedia_add(&sc->sc_ifmedia, IFM_ETHER | IFM_10_T | IFM_FDX, 0, NULL);
ifmedia_add(&sc->sc_ifmedia, IFM_ETHER | IFM_100_TX, 0, NULL);
ifmedia_add(&sc->sc_ifmedia, IFM_ETHER | IFM_100_TX | IFM_HDX, 0, NULL);
ifmedia_add(&sc->sc_ifmedia, IFM_ETHER | IFM_100_TX | IFM_FDX, 0, NULL);
ifmedia_add(&sc->sc_ifmedia, IFM_ETHER | IFM_AUTO, 0, NULL);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
pci_enable_busmaster(dev);
/* Prefer memory space register mapping over IO space. */
if (prefer_iomap == 0) {
sc->sc_res_id = PCIR_BAR(1);
sc->sc_res_type = SYS_RES_MEMORY;
} else {
sc->sc_res_id = PCIR_BAR(0);
sc->sc_res_type = SYS_RES_IOPORT;
}
sc->sc_res = bus_alloc_resource_any(dev, sc->sc_res_type,
&sc->sc_res_id, RF_ACTIVE);
if (sc->sc_res == NULL && prefer_iomap == 0) {
sc->sc_res_id = PCIR_BAR(0);
sc->sc_res_type = SYS_RES_IOPORT;
sc->sc_res = bus_alloc_resource_any(dev, sc->sc_res_type,
&sc->sc_res_id, RF_ACTIVE);
}
if (sc->sc_res == NULL) {
device_printf(dev, "couldn't map ports/memory\n");
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
ifmedia_removeall(&sc->sc_ifmedia);
mtx_destroy(&sc->sc_mtx);
return (ENXIO);
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Enable MWI. */
reg = pci_read_config(dev, PCIR_COMMAND, 2);
reg |= PCIM_CMD_MWRICEN;
pci_write_config(dev, PCIR_COMMAND, reg, 2);
/* Check cache line size. */
reg = pci_read_config(dev, PCIR_CACHELNSZ, 1);
reg <<= 4;
if (reg == 0 || (reg % 16) != 0)
device_printf(sc->sc_dev,
"invalid cache line size : %u\n", reg);
/* Allocate interrupt */
rid = 0;
sc->sc_irq = bus_alloc_resource_any(dev, SYS_RES_IRQ, &rid,
RF_SHAREABLE | RF_ACTIVE);
if (sc->sc_irq == NULL) {
device_printf(dev, "couldn't map interrupt\n");
error = ENXIO;
goto fail;
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if ((error = txp_alloc_rings(sc)) != 0)
goto fail;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
txp_init_rings(sc);
txp_sysctl_node(sc);
/* Reset controller and make it reload sleep image. */
if (txp_reset(sc) != 0) {
error = ENXIO;
goto fail;
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Let controller boot from sleep image. */
if (txp_boot(sc, STAT_WAITING_FOR_HOST_REQUEST) != 0) {
device_printf(sc->sc_dev, "could not boot sleep image\n");
error = ENXIO;
goto fail;
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Get station address. */
if (txp_command(sc, TXP_CMD_STATION_ADDRESS_READ, 0, 0, 0,
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
&p1, &p2, NULL, TXP_CMD_WAIT)) {
error = ENXIO;
goto fail;
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
p1 = le16toh(p1);
eaddr[0] = ((uint8_t *)&p1)[1];
eaddr[1] = ((uint8_t *)&p1)[0];
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
p2 = le32toh(p2);
eaddr[2] = ((uint8_t *)&p2)[3];
eaddr[3] = ((uint8_t *)&p2)[2];
eaddr[4] = ((uint8_t *)&p2)[1];
eaddr[5] = ((uint8_t *)&p2)[0];
ifp = sc->sc_ifp = if_alloc(IFT_ETHER);
if (ifp == NULL) {
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
device_printf(dev, "can not allocate ifnet structure\n");
error = ENOSPC;
goto fail;
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/*
* Show sleep image version information which may help to
* diagnose sleep image specific issues.
*/
rsp = NULL;
if (txp_ext_command(sc, TXP_CMD_READ_VERSION, 0, 0, 0, NULL, 0,
&rsp, TXP_CMD_WAIT)) {
device_printf(dev, "can not read sleep image version\n");
error = ENXIO;
goto fail;
}
if (rsp->rsp_numdesc == 0) {
p2 = le32toh(rsp->rsp_par2) & 0xFFFF;
device_printf(dev, "Typhoon 1.0 sleep image (2000/%02u/%02u)\n",
p2 >> 8, p2 & 0xFF);
} else if (rsp->rsp_numdesc == 2) {
p2 = le32toh(rsp->rsp_par2);
ver = (uint8_t *)(rsp + 1);
/*
* Even if datasheet says the command returns a NULL
* terminated version string, explicitly terminate
* the string. Given that several bugs of firmware
* I can't trust this simple one.
*/
ver[25] = '\0';
device_printf(dev,
"Typhoon 1.1+ sleep image %02u.%03u.%03u %s\n",
p2 >> 24, (p2 >> 12) & 0xFFF, p2 & 0xFFF, ver);
} else {
p2 = le32toh(rsp->rsp_par2);
device_printf(dev,
"Unknown Typhoon sleep image version: %u:0x%08x\n",
rsp->rsp_numdesc, p2);
}
if (rsp != NULL)
free(rsp, M_DEVBUF);
sc->sc_xcvr = TXP_XCVR_AUTO;
txp_command(sc, TXP_CMD_XCVR_SELECT, TXP_XCVR_AUTO, 0, 0,
NULL, NULL, NULL, TXP_CMD_NOWAIT);
ifmedia_set(&sc->sc_ifmedia, IFM_ETHER | IFM_AUTO);
ifp->if_softc = sc;
if_initname(ifp, device_get_name(dev), device_get_unit(dev));
ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
ifp->if_ioctl = txp_ioctl;
ifp->if_start = txp_start;
ifp->if_init = txp_init;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
ifp->if_snd.ifq_drv_maxlen = TX_ENTRIES - 1;
IFQ_SET_MAXLEN(&ifp->if_snd, ifp->if_snd.ifq_drv_maxlen);
IFQ_SET_READY(&ifp->if_snd);
/*
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
* It's possible to read firmware's offload capability but
* we have not downloaded the firmware yet so announce
* working capability here. We're not interested in IPSec
* capability and due to the lots of firmware bug we can't
* advertise the whole capability anyway.
*/
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
ifp->if_capabilities = IFCAP_RXCSUM | IFCAP_TXCSUM;
if (pci_find_cap(dev, PCIY_PMG, &pmc) == 0)
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
ifp->if_capabilities |= IFCAP_WOL_MAGIC;
/* Enable all capabilities. */
ifp->if_capenable = ifp->if_capabilities;
ether_ifattach(ifp, eaddr);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* VLAN capability setup. */
ifp->if_capabilities |= IFCAP_VLAN_MTU;
ifp->if_capabilities |= IFCAP_VLAN_HWTAGGING | IFCAP_VLAN_HWCSUM;
ifp->if_capenable = ifp->if_capabilities;
/* Tell the upper layer(s) we support long frames. */
ifp->if_data.ifi_hdrlen = sizeof(struct ether_vlan_header);
WRITE_REG(sc, TXP_IER, TXP_INTR_NONE);
WRITE_REG(sc, TXP_IMR, TXP_INTR_ALL);
/* Create local taskq. */
sc->sc_tq = taskqueue_create_fast("txp_taskq", M_WAITOK,
taskqueue_thread_enqueue, &sc->sc_tq);
if (sc->sc_tq == NULL) {
device_printf(dev, "could not create taskqueue.\n");
ether_ifdetach(ifp);
error = ENXIO;
goto fail;
}
taskqueue_start_threads(&sc->sc_tq, 1, PI_NET, "%s taskq",
device_get_nameunit(sc->sc_dev));
/* Put controller into sleep. */
if (txp_sleep(sc, 0) != 0) {
ether_ifdetach(ifp);
error = ENXIO;
goto fail;
}
error = bus_setup_intr(dev, sc->sc_irq, INTR_TYPE_NET | INTR_MPSAFE,
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
txp_intr, NULL, sc, &sc->sc_intrhand);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if (error != 0) {
ether_ifdetach(ifp);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
device_printf(dev, "couldn't set up interrupt handler.\n");
goto fail;
}
2009-03-11 08:49:17 +00:00
return (0);
fail:
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if (error != 0)
txp_detach(dev);
2009-03-11 08:49:17 +00:00
return (error);
}
static int
2009-03-11 08:25:18 +00:00
txp_detach(device_t dev)
{
struct txp_softc *sc;
struct ifnet *ifp;
sc = device_get_softc(dev);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
ifp = sc->sc_ifp;
if (device_is_attached(dev)) {
TXP_LOCK(sc);
sc->sc_flags |= TXP_FLAG_DETACH;
txp_stop(sc);
TXP_UNLOCK(sc);
callout_drain(&sc->sc_tick);
taskqueue_drain(sc->sc_tq, &sc->sc_int_task);
ether_ifdetach(ifp);
}
WRITE_REG(sc, TXP_IMR, TXP_INTR_ALL);
ifmedia_removeall(&sc->sc_ifmedia);
if (sc->sc_intrhand != NULL)
bus_teardown_intr(dev, sc->sc_irq, sc->sc_intrhand);
if (sc->sc_irq != NULL)
bus_release_resource(dev, SYS_RES_IRQ, 0, sc->sc_irq);
if (sc->sc_res != NULL)
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
bus_release_resource(dev, sc->sc_res_type, sc->sc_res_id,
sc->sc_res);
if (sc->sc_ifp != NULL) {
if_free(sc->sc_ifp);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
sc->sc_ifp = NULL;
}
txp_free_rings(sc);
mtx_destroy(&sc->sc_mtx);
return (0);
}
static int
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
txp_reset(struct txp_softc *sc)
{
uint32_t r;
int i;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Disable interrupts. */
WRITE_REG(sc, TXP_IER, TXP_INTR_NONE);
WRITE_REG(sc, TXP_IMR, TXP_INTR_ALL);
/* Ack all pending interrupts. */
WRITE_REG(sc, TXP_ISR, TXP_INTR_ALL);
r = 0;
WRITE_REG(sc, TXP_SRR, TXP_SRR_ALL);
DELAY(1000);
WRITE_REG(sc, TXP_SRR, 0);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Should wait max 6 seconds. */
for (i = 0; i < 6000; i++) {
r = READ_REG(sc, TXP_A2H_0);
if (r == STAT_WAITING_FOR_HOST_REQUEST)
break;
DELAY(1000);
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if (r != STAT_WAITING_FOR_HOST_REQUEST)
device_printf(sc->sc_dev, "reset hung\n");
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
WRITE_REG(sc, TXP_IER, TXP_INTR_NONE);
WRITE_REG(sc, TXP_IMR, TXP_INTR_ALL);
WRITE_REG(sc, TXP_ISR, TXP_INTR_ALL);
/*
* Give more time to complete loading sleep image before
* trying to boot from sleep image.
*/
DELAY(5000);
return (0);
}
static int
txp_boot(struct txp_softc *sc, uint32_t state)
{
/* See if it's waiting for boot, and try to boot it. */
if (txp_wait(sc, state) != 0) {
device_printf(sc->sc_dev, "not waiting for boot\n");
return (ENXIO);
}
WRITE_REG(sc, TXP_H2A_2, TXP_ADDR_HI(sc->sc_ldata.txp_boot_paddr));
TXP_BARRIER(sc, TXP_H2A_2, 4, BUS_SPACE_BARRIER_WRITE);
WRITE_REG(sc, TXP_H2A_1, TXP_ADDR_LO(sc->sc_ldata.txp_boot_paddr));
TXP_BARRIER(sc, TXP_H2A_1, 4, BUS_SPACE_BARRIER_WRITE);
WRITE_REG(sc, TXP_H2A_0, TXP_BOOTCMD_REGISTER_BOOT_RECORD);
TXP_BARRIER(sc, TXP_H2A_0, 4, BUS_SPACE_BARRIER_WRITE);
/* See if it booted. */
if (txp_wait(sc, STAT_RUNNING) != 0) {
device_printf(sc->sc_dev, "firmware not running\n");
return (ENXIO);
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Clear TX and CMD ring write registers. */
WRITE_REG(sc, TXP_H2A_1, TXP_BOOTCMD_NULL);
TXP_BARRIER(sc, TXP_H2A_1, 4, BUS_SPACE_BARRIER_WRITE);
WRITE_REG(sc, TXP_H2A_2, TXP_BOOTCMD_NULL);
TXP_BARRIER(sc, TXP_H2A_2, 4, BUS_SPACE_BARRIER_WRITE);
WRITE_REG(sc, TXP_H2A_3, TXP_BOOTCMD_NULL);
TXP_BARRIER(sc, TXP_H2A_3, 4, BUS_SPACE_BARRIER_WRITE);
WRITE_REG(sc, TXP_H2A_0, TXP_BOOTCMD_NULL);
TXP_BARRIER(sc, TXP_H2A_0, 4, BUS_SPACE_BARRIER_WRITE);
return (0);
}
static int
2009-03-11 08:25:18 +00:00
txp_download_fw(struct txp_softc *sc)
{
struct txp_fw_file_header *fileheader;
struct txp_fw_section_header *secthead;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
int sect;
uint32_t error, ier, imr;
TXP_LOCK_ASSERT(sc);
error = 0;
ier = READ_REG(sc, TXP_IER);
WRITE_REG(sc, TXP_IER, ier | TXP_INT_A2H_0);
imr = READ_REG(sc, TXP_IMR);
WRITE_REG(sc, TXP_IMR, imr | TXP_INT_A2H_0);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if (txp_wait(sc, STAT_WAITING_FOR_HOST_REQUEST) != 0) {
device_printf(sc->sc_dev, "not waiting for host request\n");
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
error = ETIMEDOUT;
goto fail;
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Ack the status. */
WRITE_REG(sc, TXP_ISR, TXP_INT_A2H_0);
fileheader = (struct txp_fw_file_header *)tc990image;
if (bcmp("TYPHOON", fileheader->magicid, sizeof(fileheader->magicid))) {
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
device_printf(sc->sc_dev, "firmware invalid magic\n");
goto fail;
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Tell boot firmware to get ready for image. */
WRITE_REG(sc, TXP_H2A_1, le32toh(fileheader->addr));
TXP_BARRIER(sc, TXP_H2A_1, 4, BUS_SPACE_BARRIER_WRITE);
WRITE_REG(sc, TXP_H2A_2, le32toh(fileheader->hmac[0]));
TXP_BARRIER(sc, TXP_H2A_2, 4, BUS_SPACE_BARRIER_WRITE);
WRITE_REG(sc, TXP_H2A_3, le32toh(fileheader->hmac[1]));
TXP_BARRIER(sc, TXP_H2A_3, 4, BUS_SPACE_BARRIER_WRITE);
WRITE_REG(sc, TXP_H2A_4, le32toh(fileheader->hmac[2]));
TXP_BARRIER(sc, TXP_H2A_4, 4, BUS_SPACE_BARRIER_WRITE);
WRITE_REG(sc, TXP_H2A_5, le32toh(fileheader->hmac[3]));
TXP_BARRIER(sc, TXP_H2A_5, 4, BUS_SPACE_BARRIER_WRITE);
WRITE_REG(sc, TXP_H2A_6, le32toh(fileheader->hmac[4]));
TXP_BARRIER(sc, TXP_H2A_6, 4, BUS_SPACE_BARRIER_WRITE);
WRITE_REG(sc, TXP_H2A_0, TXP_BOOTCMD_RUNTIME_IMAGE);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
TXP_BARRIER(sc, TXP_H2A_0, 4, BUS_SPACE_BARRIER_WRITE);
if (txp_download_fw_wait(sc)) {
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
device_printf(sc->sc_dev, "firmware wait failed, initial\n");
error = ETIMEDOUT;
goto fail;
}
secthead = (struct txp_fw_section_header *)(((uint8_t *)tc990image) +
sizeof(struct txp_fw_file_header));
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
for (sect = 0; sect < le32toh(fileheader->nsections); sect++) {
if ((error = txp_download_fw_section(sc, secthead, sect)) != 0)
goto fail;
secthead = (struct txp_fw_section_header *)
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
(((uint8_t *)secthead) + le32toh(secthead->nbytes) +
sizeof(*secthead));
}
WRITE_REG(sc, TXP_H2A_0, TXP_BOOTCMD_DOWNLOAD_COMPLETE);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
TXP_BARRIER(sc, TXP_H2A_0, 4, BUS_SPACE_BARRIER_WRITE);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if (txp_wait(sc, STAT_WAITING_FOR_BOOT) != 0) {
device_printf(sc->sc_dev, "not waiting for boot\n");
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
error = ETIMEDOUT;
goto fail;
}
fail:
WRITE_REG(sc, TXP_IER, ier);
WRITE_REG(sc, TXP_IMR, imr);
return (error);
}
static int
2009-03-11 08:25:18 +00:00
txp_download_fw_wait(struct txp_softc *sc)
{
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
uint32_t i;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
TXP_LOCK_ASSERT(sc);
for (i = 0; i < TXP_TIMEOUT; i++) {
if ((READ_REG(sc, TXP_ISR) & TXP_INT_A2H_0) != 0)
break;
DELAY(50);
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if (i == TXP_TIMEOUT) {
device_printf(sc->sc_dev, "firmware wait failed comm0\n");
return (ETIMEDOUT);
}
WRITE_REG(sc, TXP_ISR, TXP_INT_A2H_0);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if (READ_REG(sc, TXP_A2H_0) != STAT_WAITING_FOR_SEGMENT) {
device_printf(sc->sc_dev, "firmware not waiting for segment\n");
return (ETIMEDOUT);
}
return (0);
}
static int
2009-03-11 08:25:18 +00:00
txp_download_fw_section(struct txp_softc *sc,
struct txp_fw_section_header *sect, int sectnum)
{
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
bus_dma_tag_t sec_tag;
bus_dmamap_t sec_map;
bus_addr_t sec_paddr;
uint8_t *sec_buf;
int rseg, err = 0;
struct mbuf m;
uint16_t csum;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
TXP_LOCK_ASSERT(sc);
/* Skip zero length sections. */
if (le32toh(sect->nbytes) == 0)
return (0);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Make sure we aren't past the end of the image. */
rseg = ((uint8_t *)sect) - ((uint8_t *)tc990image);
if (rseg >= sizeof(tc990image)) {
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
device_printf(sc->sc_dev,
"firmware invalid section address, section %d\n", sectnum);
return (EIO);
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Make sure this section doesn't go past the end. */
rseg += le32toh(sect->nbytes);
if (rseg >= sizeof(tc990image)) {
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
device_printf(sc->sc_dev, "firmware truncated section %d\n",
sectnum);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
return (EIO);
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
sec_tag = NULL;
sec_map = NULL;
sec_buf = NULL;
/* XXX */
TXP_UNLOCK(sc);
err = txp_dma_alloc(sc, "firmware sections", &sec_tag, sizeof(uint32_t),
0, &sec_map, (void **)&sec_buf, le32toh(sect->nbytes), &sec_paddr);
TXP_LOCK(sc);
if (err != 0)
goto bail;
bcopy(((uint8_t *)sect) + sizeof(*sect), sec_buf,
le32toh(sect->nbytes));
/*
* dummy up mbuf and verify section checksum
*/
m.m_type = MT_DATA;
m.m_next = m.m_nextpkt = NULL;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
m.m_len = le32toh(sect->nbytes);
m.m_data = sec_buf;
m.m_flags = 0;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
csum = in_cksum(&m, le32toh(sect->nbytes));
if (csum != sect->cksum) {
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
device_printf(sc->sc_dev,
"firmware section %d, bad cksum (expected 0x%x got 0x%x)\n",
sectnum, le16toh(sect->cksum), csum);
err = EIO;
goto bail;
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
bus_dmamap_sync(sec_tag, sec_map, BUS_DMASYNC_PREWRITE);
WRITE_REG(sc, TXP_H2A_1, le32toh(sect->nbytes));
TXP_BARRIER(sc, TXP_H2A_1, 4, BUS_SPACE_BARRIER_WRITE);
WRITE_REG(sc, TXP_H2A_2, le16toh(sect->cksum));
TXP_BARRIER(sc, TXP_H2A_2, 4, BUS_SPACE_BARRIER_WRITE);
WRITE_REG(sc, TXP_H2A_3, le32toh(sect->addr));
TXP_BARRIER(sc, TXP_H2A_3, 4, BUS_SPACE_BARRIER_WRITE);
WRITE_REG(sc, TXP_H2A_4, TXP_ADDR_HI(sec_paddr));
TXP_BARRIER(sc, TXP_H2A_4, 4, BUS_SPACE_BARRIER_WRITE);
WRITE_REG(sc, TXP_H2A_5, TXP_ADDR_LO(sec_paddr));
TXP_BARRIER(sc, TXP_H2A_5, 4, BUS_SPACE_BARRIER_WRITE);
WRITE_REG(sc, TXP_H2A_0, TXP_BOOTCMD_SEGMENT_AVAILABLE);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
TXP_BARRIER(sc, TXP_H2A_0, 4, BUS_SPACE_BARRIER_WRITE);
if (txp_download_fw_wait(sc)) {
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
device_printf(sc->sc_dev,
"firmware wait failed, section %d\n", sectnum);
err = ETIMEDOUT;
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
bus_dmamap_sync(sec_tag, sec_map, BUS_DMASYNC_POSTWRITE);
bail:
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
txp_dma_free(sc, &sec_tag, &sec_map, (void **)&sec_buf);
return (err);
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
static int
2009-03-11 08:25:18 +00:00
txp_intr(void *vsc)
{
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
struct txp_softc *sc;
uint32_t status;
sc = vsc;
status = READ_REG(sc, TXP_ISR);
if ((status & TXP_INT_LATCH) == 0)
return (FILTER_STRAY);
WRITE_REG(sc, TXP_ISR, status);
WRITE_REG(sc, TXP_IMR, TXP_INTR_ALL);
taskqueue_enqueue(sc->sc_tq, &sc->sc_int_task);
return (FILTER_HANDLED);
}
static void
txp_int_task(void *arg, int pending)
{
struct txp_softc *sc;
struct ifnet *ifp;
struct txp_hostvar *hv;
uint32_t isr;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
int more;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
sc = (struct txp_softc *)arg;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
TXP_LOCK(sc);
ifp = sc->sc_ifp;
hv = sc->sc_hostvar;
isr = READ_REG(sc, TXP_ISR);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if ((isr & TXP_INT_LATCH) != 0)
WRITE_REG(sc, TXP_ISR, isr);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if ((ifp->if_drv_flags & IFF_DRV_RUNNING) != 0) {
bus_dmamap_sync(sc->sc_cdata.txp_hostvar_tag,
sc->sc_cdata.txp_hostvar_map,
BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
more = 0;
if ((*sc->sc_rxhir.r_roff) != (*sc->sc_rxhir.r_woff))
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
more += txp_rx_reclaim(sc, &sc->sc_rxhir,
sc->sc_process_limit);
if ((*sc->sc_rxlor.r_roff) != (*sc->sc_rxlor.r_woff))
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
more += txp_rx_reclaim(sc, &sc->sc_rxlor,
sc->sc_process_limit);
/*
* XXX
* It seems controller is not smart enough to handle
* FIFO overflow conditions under heavy network load.
* No matter how often new Rx buffers are passed to
* controller the situation didn't change. Maybe
* flow-control would be the only way to mitigate the
* issue but firmware does not have commands that
* control the threshold of emitting pause frames.
*/
if (hv->hv_rx_buf_write_idx == hv->hv_rx_buf_read_idx)
txp_rxbuf_reclaim(sc);
if (sc->sc_txhir.r_cnt && (sc->sc_txhir.r_cons !=
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
TXP_OFFSET2IDX(le32toh(*(sc->sc_txhir.r_off)))))
txp_tx_reclaim(sc, &sc->sc_txhir);
if (sc->sc_txlor.r_cnt && (sc->sc_txlor.r_cons !=
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
TXP_OFFSET2IDX(le32toh(*(sc->sc_txlor.r_off)))))
txp_tx_reclaim(sc, &sc->sc_txlor);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
bus_dmamap_sync(sc->sc_cdata.txp_hostvar_tag,
sc->sc_cdata.txp_hostvar_map,
BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
if (!IFQ_DRV_IS_EMPTY(&ifp->if_snd))
txp_start_locked(sc->sc_ifp);
if (more != 0 || READ_REG(sc, TXP_ISR & TXP_INT_LATCH) != 0) {
taskqueue_enqueue(sc->sc_tq, &sc->sc_int_task);
TXP_UNLOCK(sc);
return;
}
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Re-enable interrupts. */
WRITE_REG(sc, TXP_IMR, TXP_INTR_NONE);
TXP_UNLOCK(sc);
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
#ifndef __NO_STRICT_ALIGNMENT
static __inline void
txp_fixup_rx(struct mbuf *m)
{
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
int i;
uint16_t *src, *dst;
src = mtod(m, uint16_t *);
dst = src - (TXP_RXBUF_ALIGN - ETHER_ALIGN) / sizeof *src;
for (i = 0; i < (m->m_len / sizeof(uint16_t) + 1); i++)
*dst++ = *src++;
m->m_data -= TXP_RXBUF_ALIGN - ETHER_ALIGN;
}
#endif
static int
txp_rx_reclaim(struct txp_softc *sc, struct txp_rx_ring *r, int count)
{
struct ifnet *ifp;
struct txp_rx_desc *rxd;
struct mbuf *m;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
struct txp_rx_swdesc *sd;
uint32_t roff, woff, rx_stat, prog;
TXP_LOCK_ASSERT(sc);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
ifp = sc->sc_ifp;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
bus_dmamap_sync(r->r_tag, r->r_map, BUS_DMASYNC_POSTREAD |
BUS_DMASYNC_POSTWRITE);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
roff = le32toh(*r->r_roff);
woff = le32toh(*r->r_woff);
rxd = r->r_desc + roff / sizeof(struct txp_rx_desc);
for (prog = 0; roff != woff; prog++, count--) {
if (count <= 0)
break;
bcopy((u_long *)&rxd->rx_vaddrlo, &sd, sizeof(sd));
KASSERT(sd != NULL, ("%s: Rx desc ring corrupted", __func__));
bus_dmamap_sync(sc->sc_cdata.txp_rx_tag, sd->sd_map,
BUS_DMASYNC_POSTREAD);
bus_dmamap_unload(sc->sc_cdata.txp_rx_tag, sd->sd_map);
m = sd->sd_mbuf;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
KASSERT(m != NULL, ("%s: Rx buffer ring corrupted", __func__));
sd->sd_mbuf = NULL;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
TAILQ_REMOVE(&sc->sc_busy_list, sd, sd_next);
TAILQ_INSERT_TAIL(&sc->sc_free_list, sd, sd_next);
if ((rxd->rx_flags & RX_FLAGS_ERROR) != 0) {
if (bootverbose)
device_printf(sc->sc_dev, "Rx error %u\n",
le32toh(rxd->rx_stat) & RX_ERROR_MASK);
m_freem(m);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
goto next;
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
m->m_pkthdr.len = m->m_len = le16toh(rxd->rx_len);
m->m_pkthdr.rcvif = ifp;
#ifndef __NO_STRICT_ALIGNMENT
txp_fixup_rx(m);
#endif
rx_stat = le32toh(rxd->rx_stat);
if ((ifp->if_capenable & IFCAP_RXCSUM) != 0) {
if ((rx_stat & RX_STAT_IPCKSUMBAD) != 0)
m->m_pkthdr.csum_flags |= CSUM_IP_CHECKED;
else if ((rx_stat & RX_STAT_IPCKSUMGOOD) != 0)
m->m_pkthdr.csum_flags |=
CSUM_IP_CHECKED|CSUM_IP_VALID;
if ((rx_stat & RX_STAT_TCPCKSUMGOOD) != 0 ||
(rx_stat & RX_STAT_UDPCKSUMGOOD) != 0) {
m->m_pkthdr.csum_flags |=
CSUM_DATA_VALID | CSUM_PSEUDO_HDR;
m->m_pkthdr.csum_data = 0xffff;
}
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/*
* XXX
* Typhoon has a firmware bug that VLAN tag is always
* stripped out even if it is told to not remove the tag.
* Therefore don't check if_capenable here.
*/
if (/* (ifp->if_capenable & IFCAP_VLAN_HWTAGGING) != 0 && */
(rx_stat & RX_STAT_VLAN) != 0) {
m->m_pkthdr.ether_vtag =
bswap16((le32toh(rxd->rx_vlan) >> 16));
m->m_flags |= M_VLANTAG;
}
TXP_UNLOCK(sc);
(*ifp->if_input)(ifp, m);
TXP_LOCK(sc);
next:
roff += sizeof(struct txp_rx_desc);
if (roff == (RX_ENTRIES * sizeof(struct txp_rx_desc))) {
roff = 0;
rxd = r->r_desc;
} else
rxd++;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
prog++;
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if (prog == 0)
return (0);
bus_dmamap_sync(r->r_tag, r->r_map, BUS_DMASYNC_PREREAD |
BUS_DMASYNC_PREWRITE);
*r->r_roff = le32toh(roff);
return (count > 0 ? 0 : EAGAIN);
}
static void
2009-03-11 08:25:18 +00:00
txp_rxbuf_reclaim(struct txp_softc *sc)
{
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
struct txp_hostvar *hv;
struct txp_rxbuf_desc *rbd;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
struct txp_rx_swdesc *sd;
bus_dma_segment_t segs[1];
int nsegs, prod, prog;
uint32_t cons;
TXP_LOCK_ASSERT(sc);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
hv = sc->sc_hostvar;
cons = TXP_OFFSET2IDX(le32toh(hv->hv_rx_buf_read_idx));
prod = sc->sc_rxbufprod;
TXP_DESC_INC(prod, RXBUF_ENTRIES);
if (prod == cons)
return;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
bus_dmamap_sync(sc->sc_cdata.txp_rxbufs_tag,
sc->sc_cdata.txp_rxbufs_map,
BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
for (prog = 0; prod != cons; prog++) {
sd = TAILQ_FIRST(&sc->sc_free_list);
if (sd == NULL)
break;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
rbd = sc->sc_rxbufs + prod;
bcopy((u_long *)&rbd->rb_vaddrlo, &sd, sizeof(sd));
sd->sd_mbuf = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
if (sd->sd_mbuf == NULL)
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
break;
sd->sd_mbuf->m_pkthdr.len = sd->sd_mbuf->m_len = MCLBYTES;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
#ifndef __NO_STRICT_ALIGNMENT
m_adj(sd->sd_mbuf, TXP_RXBUF_ALIGN);
#endif
if (bus_dmamap_load_mbuf_sg(sc->sc_cdata.txp_rx_tag,
sd->sd_map, sd->sd_mbuf, segs, &nsegs, 0) != 0) {
m_freem(sd->sd_mbuf);
sd->sd_mbuf = NULL;
break;
}
KASSERT(nsegs == 1, ("%s : %d segments returned!", __func__,
nsegs));
TAILQ_REMOVE(&sc->sc_free_list, sd, sd_next);
TAILQ_INSERT_TAIL(&sc->sc_busy_list, sd, sd_next);
bus_dmamap_sync(sc->sc_cdata.txp_rx_tag, sd->sd_map,
BUS_DMASYNC_PREREAD);
rbd->rb_paddrlo = htole32(TXP_ADDR_LO(segs[0].ds_addr));
rbd->rb_paddrhi = htole32(TXP_ADDR_HI(segs[0].ds_addr));
TXP_DESC_INC(prod, RXBUF_ENTRIES);
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if (prog == 0)
return;
bus_dmamap_sync(sc->sc_cdata.txp_rxbufs_tag,
sc->sc_cdata.txp_rxbufs_map,
BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
prod = (prod + RXBUF_ENTRIES - 1) % RXBUF_ENTRIES;
sc->sc_rxbufprod = prod;
hv->hv_rx_buf_write_idx = htole32(TXP_IDX2OFFSET(prod));
}
/*
* Reclaim mbufs and entries from a transmit ring.
*/
static void
2009-03-11 08:25:18 +00:00
txp_tx_reclaim(struct txp_softc *sc, struct txp_tx_ring *r)
{
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
struct ifnet *ifp;
uint32_t idx;
uint32_t cons, cnt;
struct txp_tx_desc *txd;
struct txp_swdesc *sd;
TXP_LOCK_ASSERT(sc);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
bus_dmamap_sync(r->r_tag, r->r_map, BUS_DMASYNC_POSTREAD |
BUS_DMASYNC_POSTWRITE);
ifp = sc->sc_ifp;
idx = TXP_OFFSET2IDX(le32toh(*(r->r_off)));
cons = r->r_cons;
cnt = r->r_cnt;
txd = r->r_desc + cons;
sd = sc->sc_txd + cons;
for (cnt = r->r_cnt; cons != idx && cnt > 0; cnt--) {
if ((txd->tx_flags & TX_FLAGS_TYPE_M) == TX_FLAGS_TYPE_DATA) {
if (sd->sd_mbuf != NULL) {
bus_dmamap_sync(sc->sc_cdata.txp_tx_tag,
sd->sd_map, BUS_DMASYNC_POSTWRITE);
bus_dmamap_unload(sc->sc_cdata.txp_tx_tag,
sd->sd_map);
m_freem(sd->sd_mbuf);
sd->sd_mbuf = NULL;
txd->tx_addrlo = 0;
txd->tx_addrhi = 0;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
txd->tx_flags = 0;
}
}
ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
if (++cons == TX_ENTRIES) {
txd = r->r_desc;
cons = 0;
sd = sc->sc_txd;
} else {
txd++;
sd++;
}
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
bus_dmamap_sync(r->r_tag, r->r_map, BUS_DMASYNC_PREREAD |
BUS_DMASYNC_PREWRITE);
r->r_cons = cons;
r->r_cnt = cnt;
if (cnt == 0)
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
sc->sc_watchdog_timer = 0;
}
static int
2009-03-11 08:25:18 +00:00
txp_shutdown(device_t dev)
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
{
return (txp_suspend(dev));
}
static int
txp_suspend(device_t dev)
{
struct txp_softc *sc;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
struct ifnet *ifp;
uint8_t *eaddr;
uint16_t p1;
uint32_t p2;
int pmc;
uint16_t pmstat;
sc = device_get_softc(dev);
TXP_LOCK(sc);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
ifp = sc->sc_ifp;
txp_stop(sc);
txp_init_rings(sc);
/* Reset controller and make it reload sleep image. */
txp_reset(sc);
/* Let controller boot from sleep image. */
if (txp_boot(sc, STAT_WAITING_FOR_HOST_REQUEST) != 0)
device_printf(sc->sc_dev, "couldn't boot sleep image\n");
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Set station address. */
eaddr = IF_LLADDR(sc->sc_ifp);
p1 = 0;
((uint8_t *)&p1)[1] = eaddr[0];
((uint8_t *)&p1)[0] = eaddr[1];
p1 = le16toh(p1);
((uint8_t *)&p2)[3] = eaddr[2];
((uint8_t *)&p2)[2] = eaddr[3];
((uint8_t *)&p2)[1] = eaddr[4];
((uint8_t *)&p2)[0] = eaddr[5];
p2 = le32toh(p2);
txp_command(sc, TXP_CMD_STATION_ADDRESS_WRITE, p1, p2, 0, NULL, NULL,
NULL, TXP_CMD_WAIT);
txp_set_filter(sc);
WRITE_REG(sc, TXP_IER, TXP_INTR_NONE);
WRITE_REG(sc, TXP_IMR, TXP_INTR_ALL);
txp_sleep(sc, sc->sc_ifp->if_capenable);
if (pci_find_cap(sc->sc_dev, PCIY_PMG, &pmc) == 0) {
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Request PME. */
pmstat = pci_read_config(sc->sc_dev,
pmc + PCIR_POWER_STATUS, 2);
pmstat &= ~(PCIM_PSTAT_PME | PCIM_PSTAT_PMEENABLE);
if ((ifp->if_capenable & IFCAP_WOL) != 0)
pmstat |= PCIM_PSTAT_PME | PCIM_PSTAT_PMEENABLE;
pci_write_config(sc->sc_dev,
pmc + PCIR_POWER_STATUS, pmstat, 2);
}
TXP_UNLOCK(sc);
return (0);
}
static int
txp_resume(device_t dev)
{
struct txp_softc *sc;
int pmc;
uint16_t pmstat;
sc = device_get_softc(dev);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
TXP_LOCK(sc);
if (pci_find_cap(sc->sc_dev, PCIY_PMG, &pmc) == 0) {
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Disable PME and clear PME status. */
pmstat = pci_read_config(sc->sc_dev,
pmc + PCIR_POWER_STATUS, 2);
if ((pmstat & PCIM_PSTAT_PMEENABLE) != 0) {
pmstat &= ~PCIM_PSTAT_PMEENABLE;
pci_write_config(sc->sc_dev,
pmc + PCIR_POWER_STATUS, pmstat, 2);
}
}
if ((sc->sc_ifp->if_flags & IFF_UP) != 0)
txp_init_locked(sc);
TXP_UNLOCK(sc);
2009-03-11 08:49:17 +00:00
return (0);
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
struct txp_dmamap_arg {
bus_addr_t txp_busaddr;
};
static void
txp_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
{
struct txp_dmamap_arg *ctx;
if (error != 0)
return;
KASSERT(nsegs == 1, ("%s: %d segments returned!", __func__, nsegs));
ctx = (struct txp_dmamap_arg *)arg;
ctx->txp_busaddr = segs[0].ds_addr;
}
static int
txp_dma_alloc(struct txp_softc *sc, char *type, bus_dma_tag_t *tag,
bus_size_t alignment, bus_size_t boundary, bus_dmamap_t *map, void **buf,
bus_size_t size, bus_addr_t *paddr)
{
struct txp_dmamap_arg ctx;
int error;
/* Create DMA block tag. */
error = bus_dma_tag_create(
sc->sc_cdata.txp_parent_tag, /* parent */
alignment, boundary, /* algnmnt, boundary */
BUS_SPACE_MAXADDR, /* lowaddr */
BUS_SPACE_MAXADDR, /* highaddr */
NULL, NULL, /* filter, filterarg */
size, /* maxsize */
1, /* nsegments */
size, /* maxsegsize */
0, /* flags */
NULL, NULL, /* lockfunc, lockarg */
tag);
if (error != 0) {
device_printf(sc->sc_dev,
"could not create DMA tag for %s.\n", type);
return (error);
}
*paddr = 0;
/* Allocate DMA'able memory and load the DMA map. */
error = bus_dmamem_alloc(*tag, buf, BUS_DMA_WAITOK | BUS_DMA_ZERO |
BUS_DMA_COHERENT, map);
if (error != 0) {
device_printf(sc->sc_dev,
"could not allocate DMA'able memory for %s.\n", type);
return (error);
}
ctx.txp_busaddr = 0;
error = bus_dmamap_load(*tag, *map, *(uint8_t **)buf,
size, txp_dmamap_cb, &ctx, BUS_DMA_NOWAIT);
if (error != 0 || ctx.txp_busaddr == 0) {
device_printf(sc->sc_dev,
"could not load DMA'able memory for %s.\n", type);
return (error);
}
*paddr = ctx.txp_busaddr;
return (0);
}
static void
txp_dma_free(struct txp_softc *sc, bus_dma_tag_t *tag, bus_dmamap_t *map,
void **buf)
{
if (*tag != NULL) {
if (*map != NULL)
bus_dmamap_unload(*tag, *map);
if (*map != NULL && buf != NULL)
bus_dmamem_free(*tag, *(uint8_t **)buf, *map);
*(uint8_t **)buf = NULL;
*map = NULL;
bus_dma_tag_destroy(*tag);
*tag = NULL;
}
}
static int
2009-03-11 08:25:18 +00:00
txp_alloc_rings(struct txp_softc *sc)
{
struct txp_boot_record *boot;
struct txp_ldata *ld;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
struct txp_swdesc *txd;
struct txp_rxbuf_desc *rbd;
struct txp_rx_swdesc *sd;
int error, i;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
ld = &sc->sc_ldata;
boot = ld->txp_boot;
/* boot record */
sc->sc_boot = boot;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/*
* Create parent ring/DMA block tag.
* Datasheet says that all ring addresses and descriptors
* support 64bits addressing. However the controller is
* known to have no support DAC so limit DMA address space
* to 32bits.
*/
error = bus_dma_tag_create(
bus_get_dma_tag(sc->sc_dev), /* parent */
1, 0, /* algnmnt, boundary */
BUS_SPACE_MAXADDR_32BIT, /* lowaddr */
BUS_SPACE_MAXADDR, /* highaddr */
NULL, NULL, /* filter, filterarg */
BUS_SPACE_MAXSIZE_32BIT, /* maxsize */
0, /* nsegments */
BUS_SPACE_MAXSIZE_32BIT, /* maxsegsize */
0, /* flags */
NULL, NULL, /* lockfunc, lockarg */
&sc->sc_cdata.txp_parent_tag);
if (error != 0) {
device_printf(sc->sc_dev, "could not create parent DMA tag.\n");
return (error);
}
/* Boot record. */
error = txp_dma_alloc(sc, "boot record",
&sc->sc_cdata.txp_boot_tag, sizeof(uint32_t), 0,
&sc->sc_cdata.txp_boot_map, (void **)&sc->sc_ldata.txp_boot,
sizeof(struct txp_boot_record),
&sc->sc_ldata.txp_boot_paddr);
if (error != 0)
return (error);
boot = sc->sc_ldata.txp_boot;
sc->sc_boot = boot;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Host variables. */
error = txp_dma_alloc(sc, "host variables",
&sc->sc_cdata.txp_hostvar_tag, sizeof(uint32_t), 0,
&sc->sc_cdata.txp_hostvar_map, (void **)&sc->sc_ldata.txp_hostvar,
sizeof(struct txp_hostvar),
&sc->sc_ldata.txp_hostvar_paddr);
if (error != 0)
return (error);
boot->br_hostvar_lo =
htole32(TXP_ADDR_LO(sc->sc_ldata.txp_hostvar_paddr));
boot->br_hostvar_hi =
htole32(TXP_ADDR_HI(sc->sc_ldata.txp_hostvar_paddr));
sc->sc_hostvar = sc->sc_ldata.txp_hostvar;
/* Hi priority tx ring. */
error = txp_dma_alloc(sc, "hi priority tx ring",
&sc->sc_cdata.txp_txhiring_tag, sizeof(struct txp_tx_desc), 0,
&sc->sc_cdata.txp_txhiring_map, (void **)&sc->sc_ldata.txp_txhiring,
sizeof(struct txp_tx_desc) * TX_ENTRIES,
&sc->sc_ldata.txp_txhiring_paddr);
if (error != 0)
return (error);
boot->br_txhipri_lo =
htole32(TXP_ADDR_LO(sc->sc_ldata.txp_txhiring_paddr));
boot->br_txhipri_hi =
htole32(TXP_ADDR_HI(sc->sc_ldata.txp_txhiring_paddr));
boot->br_txhipri_siz =
htole32(TX_ENTRIES * sizeof(struct txp_tx_desc));
sc->sc_txhir.r_tag = sc->sc_cdata.txp_txhiring_tag;
sc->sc_txhir.r_map = sc->sc_cdata.txp_txhiring_map;
sc->sc_txhir.r_reg = TXP_H2A_1;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
sc->sc_txhir.r_desc = sc->sc_ldata.txp_txhiring;
sc->sc_txhir.r_cons = sc->sc_txhir.r_prod = sc->sc_txhir.r_cnt = 0;
sc->sc_txhir.r_off = &sc->sc_hostvar->hv_tx_hi_desc_read_idx;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Low priority tx ring. */
error = txp_dma_alloc(sc, "low priority tx ring",
&sc->sc_cdata.txp_txloring_tag, sizeof(struct txp_tx_desc), 0,
&sc->sc_cdata.txp_txloring_map, (void **)&sc->sc_ldata.txp_txloring,
sizeof(struct txp_tx_desc) * TX_ENTRIES,
&sc->sc_ldata.txp_txloring_paddr);
if (error != 0)
return (error);
boot->br_txlopri_lo =
htole32(TXP_ADDR_LO(sc->sc_ldata.txp_txloring_paddr));
boot->br_txlopri_hi =
htole32(TXP_ADDR_HI(sc->sc_ldata.txp_txloring_paddr));
boot->br_txlopri_siz =
htole32(TX_ENTRIES * sizeof(struct txp_tx_desc));
sc->sc_txlor.r_tag = sc->sc_cdata.txp_txloring_tag;
sc->sc_txlor.r_map = sc->sc_cdata.txp_txloring_map;
sc->sc_txlor.r_reg = TXP_H2A_3;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
sc->sc_txlor.r_desc = sc->sc_ldata.txp_txloring;
sc->sc_txlor.r_cons = sc->sc_txlor.r_prod = sc->sc_txlor.r_cnt = 0;
sc->sc_txlor.r_off = &sc->sc_hostvar->hv_tx_lo_desc_read_idx;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* High priority rx ring. */
error = txp_dma_alloc(sc, "hi priority rx ring",
&sc->sc_cdata.txp_rxhiring_tag,
roundup(sizeof(struct txp_rx_desc), 16), 0,
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
&sc->sc_cdata.txp_rxhiring_map, (void **)&sc->sc_ldata.txp_rxhiring,
sizeof(struct txp_rx_desc) * RX_ENTRIES,
&sc->sc_ldata.txp_rxhiring_paddr);
if (error != 0)
return (error);
boot->br_rxhipri_lo =
htole32(TXP_ADDR_LO(sc->sc_ldata.txp_rxhiring_paddr));
boot->br_rxhipri_hi =
htole32(TXP_ADDR_HI(sc->sc_ldata.txp_rxhiring_paddr));
boot->br_rxhipri_siz =
htole32(RX_ENTRIES * sizeof(struct txp_rx_desc));
sc->sc_rxhir.r_tag = sc->sc_cdata.txp_rxhiring_tag;
sc->sc_rxhir.r_map = sc->sc_cdata.txp_rxhiring_map;
sc->sc_rxhir.r_desc = sc->sc_ldata.txp_rxhiring;
sc->sc_rxhir.r_roff = &sc->sc_hostvar->hv_rx_hi_read_idx;
sc->sc_rxhir.r_woff = &sc->sc_hostvar->hv_rx_hi_write_idx;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Low priority rx ring. */
error = txp_dma_alloc(sc, "low priority rx ring",
&sc->sc_cdata.txp_rxloring_tag,
roundup(sizeof(struct txp_rx_desc), 16), 0,
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
&sc->sc_cdata.txp_rxloring_map, (void **)&sc->sc_ldata.txp_rxloring,
sizeof(struct txp_rx_desc) * RX_ENTRIES,
&sc->sc_ldata.txp_rxloring_paddr);
if (error != 0)
return (error);
boot->br_rxlopri_lo =
htole32(TXP_ADDR_LO(sc->sc_ldata.txp_rxloring_paddr));
boot->br_rxlopri_hi =
htole32(TXP_ADDR_HI(sc->sc_ldata.txp_rxloring_paddr));
boot->br_rxlopri_siz =
htole32(RX_ENTRIES * sizeof(struct txp_rx_desc));
sc->sc_rxlor.r_tag = sc->sc_cdata.txp_rxloring_tag;
sc->sc_rxlor.r_map = sc->sc_cdata.txp_rxloring_map;
sc->sc_rxlor.r_desc = sc->sc_ldata.txp_rxloring;
sc->sc_rxlor.r_roff = &sc->sc_hostvar->hv_rx_lo_read_idx;
sc->sc_rxlor.r_woff = &sc->sc_hostvar->hv_rx_lo_write_idx;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Command ring. */
error = txp_dma_alloc(sc, "command ring",
&sc->sc_cdata.txp_cmdring_tag, sizeof(struct txp_cmd_desc), 0,
&sc->sc_cdata.txp_cmdring_map, (void **)&sc->sc_ldata.txp_cmdring,
sizeof(struct txp_cmd_desc) * CMD_ENTRIES,
&sc->sc_ldata.txp_cmdring_paddr);
if (error != 0)
return (error);
boot->br_cmd_lo = htole32(TXP_ADDR_LO(sc->sc_ldata.txp_cmdring_paddr));
boot->br_cmd_hi = htole32(TXP_ADDR_HI(sc->sc_ldata.txp_cmdring_paddr));
boot->br_cmd_siz = htole32(CMD_ENTRIES * sizeof(struct txp_cmd_desc));
sc->sc_cmdring.base = sc->sc_ldata.txp_cmdring;
sc->sc_cmdring.size = CMD_ENTRIES * sizeof(struct txp_cmd_desc);
sc->sc_cmdring.lastwrite = 0;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Response ring. */
error = txp_dma_alloc(sc, "response ring",
&sc->sc_cdata.txp_rspring_tag, sizeof(struct txp_rsp_desc), 0,
&sc->sc_cdata.txp_rspring_map, (void **)&sc->sc_ldata.txp_rspring,
sizeof(struct txp_rsp_desc) * RSP_ENTRIES,
&sc->sc_ldata.txp_rspring_paddr);
if (error != 0)
return (error);
boot->br_resp_lo = htole32(TXP_ADDR_LO(sc->sc_ldata.txp_rspring_paddr));
boot->br_resp_hi = htole32(TXP_ADDR_HI(sc->sc_ldata.txp_rspring_paddr));
boot->br_resp_siz = htole32(RSP_ENTRIES * sizeof(struct txp_rsp_desc));
sc->sc_rspring.base = sc->sc_ldata.txp_rspring;
sc->sc_rspring.size = RSP_ENTRIES * sizeof(struct txp_rsp_desc);
sc->sc_rspring.lastwrite = 0;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Receive buffer ring. */
error = txp_dma_alloc(sc, "receive buffer ring",
&sc->sc_cdata.txp_rxbufs_tag, sizeof(struct txp_rxbuf_desc), 0,
&sc->sc_cdata.txp_rxbufs_map, (void **)&sc->sc_ldata.txp_rxbufs,
sizeof(struct txp_rxbuf_desc) * RXBUF_ENTRIES,
&sc->sc_ldata.txp_rxbufs_paddr);
if (error != 0)
return (error);
boot->br_rxbuf_lo =
htole32(TXP_ADDR_LO(sc->sc_ldata.txp_rxbufs_paddr));
boot->br_rxbuf_hi =
htole32(TXP_ADDR_HI(sc->sc_ldata.txp_rxbufs_paddr));
boot->br_rxbuf_siz =
htole32(RXBUF_ENTRIES * sizeof(struct txp_rxbuf_desc));
sc->sc_rxbufs = sc->sc_ldata.txp_rxbufs;
/* Zero ring. */
error = txp_dma_alloc(sc, "zero buffer",
&sc->sc_cdata.txp_zero_tag, sizeof(uint32_t), 0,
&sc->sc_cdata.txp_zero_map, (void **)&sc->sc_ldata.txp_zero,
sizeof(uint32_t), &sc->sc_ldata.txp_zero_paddr);
if (error != 0)
return (error);
boot->br_zero_lo = htole32(TXP_ADDR_LO(sc->sc_ldata.txp_zero_paddr));
boot->br_zero_hi = htole32(TXP_ADDR_HI(sc->sc_ldata.txp_zero_paddr));
bus_dmamap_sync(sc->sc_cdata.txp_boot_tag, sc->sc_cdata.txp_boot_map,
BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
/* Create Tx buffers. */
error = bus_dma_tag_create(
sc->sc_cdata.txp_parent_tag, /* parent */
1, 0, /* algnmnt, boundary */
BUS_SPACE_MAXADDR, /* lowaddr */
BUS_SPACE_MAXADDR, /* highaddr */
NULL, NULL, /* filter, filterarg */
MCLBYTES * TXP_MAXTXSEGS, /* maxsize */
TXP_MAXTXSEGS, /* nsegments */
MCLBYTES, /* maxsegsize */
0, /* flags */
NULL, NULL, /* lockfunc, lockarg */
&sc->sc_cdata.txp_tx_tag);
if (error != 0) {
device_printf(sc->sc_dev, "could not create Tx DMA tag.\n");
goto fail;
}
/* Create tag for Rx buffers. */
error = bus_dma_tag_create(
sc->sc_cdata.txp_parent_tag, /* parent */
TXP_RXBUF_ALIGN, 0, /* algnmnt, boundary */
BUS_SPACE_MAXADDR, /* lowaddr */
BUS_SPACE_MAXADDR, /* highaddr */
NULL, NULL, /* filter, filterarg */
MCLBYTES, /* maxsize */
1, /* nsegments */
MCLBYTES, /* maxsegsize */
0, /* flags */
NULL, NULL, /* lockfunc, lockarg */
&sc->sc_cdata.txp_rx_tag);
if (error != 0) {
device_printf(sc->sc_dev, "could not create Rx DMA tag.\n");
goto fail;
}
/* Create DMA maps for Tx buffers. */
for (i = 0; i < TX_ENTRIES; i++) {
txd = &sc->sc_txd[i];
txd->sd_mbuf = NULL;
txd->sd_map = NULL;
error = bus_dmamap_create(sc->sc_cdata.txp_tx_tag, 0,
&txd->sd_map);
if (error != 0) {
device_printf(sc->sc_dev,
"could not create Tx dmamap.\n");
goto fail;
}
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Create DMA maps for Rx buffers. */
for (i = 0; i < RXBUF_ENTRIES; i++) {
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
sd = malloc(sizeof(struct txp_rx_swdesc), M_DEVBUF,
M_NOWAIT | M_ZERO);
if (sd == NULL) {
error = ENOMEM;
goto fail;
}
/*
* The virtual address part of descriptor is not used
* by hardware so use that to save an ring entry. We
* need bcopy here otherwise the address wouldn't be
* valid on big-endian architectures.
*/
rbd = sc->sc_rxbufs + i;
bcopy(&sd, (u_long *)&rbd->rb_vaddrlo, sizeof(sd));
sd->sd_mbuf = NULL;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
sd->sd_map = NULL;
error = bus_dmamap_create(sc->sc_cdata.txp_rx_tag, 0,
&sd->sd_map);
if (error != 0) {
device_printf(sc->sc_dev,
"could not create Rx dmamap.\n");
goto fail;
}
TAILQ_INSERT_TAIL(&sc->sc_free_list, sd, sd_next);
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
fail:
return (error);
}
static void
txp_init_rings(struct txp_softc *sc)
{
bzero(sc->sc_ldata.txp_hostvar, sizeof(struct txp_hostvar));
bzero(sc->sc_ldata.txp_zero, sizeof(uint32_t));
sc->sc_txhir.r_cons = 0;
sc->sc_txhir.r_prod = 0;
sc->sc_txhir.r_cnt = 0;
sc->sc_txlor.r_cons = 0;
sc->sc_txlor.r_prod = 0;
sc->sc_txlor.r_cnt = 0;
sc->sc_cmdring.lastwrite = 0;
sc->sc_rspring.lastwrite = 0;
sc->sc_rxbufprod = 0;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
bus_dmamap_sync(sc->sc_cdata.txp_hostvar_tag,
sc->sc_cdata.txp_hostvar_map,
BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
static int
txp_wait(struct txp_softc *sc, uint32_t state)
{
uint32_t reg;
int i;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
for (i = 0; i < TXP_TIMEOUT; i++) {
reg = READ_REG(sc, TXP_A2H_0);
if (reg == state)
break;
DELAY(50);
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
return (i == TXP_TIMEOUT ? ETIMEDOUT : 0);
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
static void
txp_free_rings(struct txp_softc *sc)
{
struct txp_swdesc *txd;
struct txp_rx_swdesc *sd;
int i;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Tx buffers. */
if (sc->sc_cdata.txp_tx_tag != NULL) {
for (i = 0; i < TX_ENTRIES; i++) {
txd = &sc->sc_txd[i];
if (txd->sd_map != NULL) {
bus_dmamap_destroy(sc->sc_cdata.txp_tx_tag,
txd->sd_map);
txd->sd_map = NULL;
}
}
bus_dma_tag_destroy(sc->sc_cdata.txp_tx_tag);
sc->sc_cdata.txp_tx_tag = NULL;
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Rx buffers. */
if (sc->sc_cdata.txp_rx_tag != NULL) {
if (sc->sc_rxbufs != NULL) {
KASSERT(TAILQ_FIRST(&sc->sc_busy_list) == NULL,
("%s : still have busy Rx buffers", __func__));
while ((sd = TAILQ_FIRST(&sc->sc_free_list)) != NULL) {
TAILQ_REMOVE(&sc->sc_free_list, sd, sd_next);
if (sd->sd_map != NULL) {
bus_dmamap_destroy(
sc->sc_cdata.txp_rx_tag,
sd->sd_map);
sd->sd_map = NULL;
}
free(sd, M_DEVBUF);
}
}
bus_dma_tag_destroy(sc->sc_cdata.txp_rx_tag);
sc->sc_cdata.txp_rx_tag = NULL;
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Hi priority Tx ring. */
txp_dma_free(sc, &sc->sc_cdata.txp_txhiring_tag,
&sc->sc_cdata.txp_txhiring_map,
(void **)&sc->sc_ldata.txp_txhiring);
/* Low priority Tx ring. */
txp_dma_free(sc, &sc->sc_cdata.txp_txloring_tag,
&sc->sc_cdata.txp_txloring_map,
(void **)&sc->sc_ldata.txp_txloring);
/* Hi priority Rx ring. */
txp_dma_free(sc, &sc->sc_cdata.txp_rxhiring_tag,
&sc->sc_cdata.txp_rxhiring_map,
(void **)&sc->sc_ldata.txp_rxhiring);
/* Low priority Rx ring. */
txp_dma_free(sc, &sc->sc_cdata.txp_rxloring_tag,
&sc->sc_cdata.txp_rxloring_map,
(void **)&sc->sc_ldata.txp_rxloring);
/* Receive buffer ring. */
txp_dma_free(sc, &sc->sc_cdata.txp_rxbufs_tag,
&sc->sc_cdata.txp_rxbufs_map, (void **)&sc->sc_ldata.txp_rxbufs);
/* Command ring. */
txp_dma_free(sc, &sc->sc_cdata.txp_cmdring_tag,
&sc->sc_cdata.txp_cmdring_map, (void **)&sc->sc_ldata.txp_cmdring);
/* Response ring. */
txp_dma_free(sc, &sc->sc_cdata.txp_rspring_tag,
&sc->sc_cdata.txp_rspring_map, (void **)&sc->sc_ldata.txp_rspring);
/* Zero ring. */
txp_dma_free(sc, &sc->sc_cdata.txp_zero_tag,
&sc->sc_cdata.txp_zero_map, (void **)&sc->sc_ldata.txp_zero);
/* Host variables. */
txp_dma_free(sc, &sc->sc_cdata.txp_hostvar_tag,
&sc->sc_cdata.txp_hostvar_map, (void **)&sc->sc_ldata.txp_hostvar);
/* Boot record. */
txp_dma_free(sc, &sc->sc_cdata.txp_boot_tag,
&sc->sc_cdata.txp_boot_map, (void **)&sc->sc_ldata.txp_boot);
if (sc->sc_cdata.txp_parent_tag != NULL) {
bus_dma_tag_destroy(sc->sc_cdata.txp_parent_tag);
sc->sc_cdata.txp_parent_tag = NULL;
}
}
static int
2009-03-11 08:25:18 +00:00
txp_ioctl(struct ifnet *ifp, u_long command, caddr_t data)
{
struct txp_softc *sc = ifp->if_softc;
struct ifreq *ifr = (struct ifreq *)data;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
int capenable, error = 0, mask;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
switch(command) {
case SIOCSIFFLAGS:
TXP_LOCK(sc);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if ((ifp->if_flags & IFF_UP) != 0) {
if ((ifp->if_drv_flags & IFF_DRV_RUNNING) != 0) {
if (((ifp->if_flags ^ sc->sc_if_flags)
& (IFF_PROMISC | IFF_ALLMULTI)) != 0)
txp_set_filter(sc);
} else {
if ((sc->sc_flags & TXP_FLAG_DETACH) == 0)
txp_init_locked(sc);
}
} else {
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if ((ifp->if_drv_flags & IFF_DRV_RUNNING) != 0)
txp_stop(sc);
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
sc->sc_if_flags = ifp->if_flags;
TXP_UNLOCK(sc);
break;
case SIOCADDMULTI:
case SIOCDELMULTI:
/*
* Multicast list has changed; set the hardware
* filter accordingly.
*/
TXP_LOCK(sc);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if ((ifp->if_drv_flags & IFF_DRV_RUNNING) != 0)
txp_set_filter(sc);
TXP_UNLOCK(sc);
break;
case SIOCSIFCAP:
TXP_LOCK(sc);
capenable = ifp->if_capenable;
mask = ifr->ifr_reqcap ^ ifp->if_capenable;
if ((mask & IFCAP_TXCSUM) != 0 &&
(ifp->if_capabilities & IFCAP_TXCSUM) != 0) {
ifp->if_capenable ^= IFCAP_TXCSUM;
if ((ifp->if_capenable & IFCAP_TXCSUM) != 0)
ifp->if_hwassist |= TXP_CSUM_FEATURES;
else
ifp->if_hwassist &= ~TXP_CSUM_FEATURES;
}
if ((mask & IFCAP_RXCSUM) != 0 &&
(ifp->if_capabilities & IFCAP_RXCSUM) != 0)
ifp->if_capenable ^= IFCAP_RXCSUM;
if ((mask & IFCAP_WOL_MAGIC) != 0 &&
(ifp->if_capabilities & IFCAP_WOL_MAGIC) != 0)
ifp->if_capenable ^= IFCAP_WOL_MAGIC;
if ((mask & IFCAP_VLAN_HWTAGGING) != 0 &&
(ifp->if_capabilities & IFCAP_VLAN_HWTAGGING) != 0)
ifp->if_capenable ^= IFCAP_VLAN_HWTAGGING;
if ((mask & IFCAP_VLAN_HWCSUM) != 0 &&
(ifp->if_capabilities & IFCAP_VLAN_HWCSUM) != 0)
ifp->if_capenable ^= IFCAP_VLAN_HWCSUM;
if ((ifp->if_capenable & IFCAP_TXCSUM) == 0)
ifp->if_capenable &= ~IFCAP_VLAN_HWCSUM;
if ((ifp->if_capenable & IFCAP_VLAN_HWTAGGING) == 0)
ifp->if_capenable &= ~IFCAP_VLAN_HWCSUM;
if (capenable != ifp->if_capenable)
txp_set_capabilities(sc);
TXP_UNLOCK(sc);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
VLAN_CAPABILITIES(ifp);
break;
case SIOCGIFMEDIA:
case SIOCSIFMEDIA:
error = ifmedia_ioctl(ifp, ifr, &sc->sc_ifmedia, command);
break;
default:
error = ether_ioctl(ifp, command, data);
break;
}
2009-03-11 08:49:17 +00:00
return (error);
}
static int
2009-03-11 08:25:18 +00:00
txp_rxring_fill(struct txp_softc *sc)
{
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
struct txp_rxbuf_desc *rbd;
struct txp_rx_swdesc *sd;
bus_dma_segment_t segs[1];
int error, i, nsegs;
TXP_LOCK_ASSERT(sc);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
bus_dmamap_sync(sc->sc_cdata.txp_rxbufs_tag,
sc->sc_cdata.txp_rxbufs_map,
BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
for (i = 0; i < RXBUF_ENTRIES; i++) {
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
sd = TAILQ_FIRST(&sc->sc_free_list);
if (sd == NULL)
return (ENOMEM);
rbd = sc->sc_rxbufs + i;
bcopy(&sd, (u_long *)&rbd->rb_vaddrlo, sizeof(sd));
KASSERT(sd->sd_mbuf == NULL,
("%s : Rx buffer ring corrupted", __func__));
sd->sd_mbuf = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
if (sd->sd_mbuf == NULL)
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
return (ENOMEM);
sd->sd_mbuf->m_pkthdr.len = sd->sd_mbuf->m_len = MCLBYTES;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
#ifndef __NO_STRICT_ALIGNMENT
m_adj(sd->sd_mbuf, TXP_RXBUF_ALIGN);
#endif
if ((error = bus_dmamap_load_mbuf_sg(sc->sc_cdata.txp_rx_tag,
sd->sd_map, sd->sd_mbuf, segs, &nsegs, 0)) != 0) {
m_freem(sd->sd_mbuf);
sd->sd_mbuf = NULL;
return (error);
}
KASSERT(nsegs == 1, ("%s : %d segments returned!", __func__,
nsegs));
TAILQ_REMOVE(&sc->sc_free_list, sd, sd_next);
TAILQ_INSERT_TAIL(&sc->sc_busy_list, sd, sd_next);
bus_dmamap_sync(sc->sc_cdata.txp_rx_tag, sd->sd_map,
BUS_DMASYNC_PREREAD);
rbd->rb_paddrlo = htole32(TXP_ADDR_LO(segs[0].ds_addr));
rbd->rb_paddrhi = htole32(TXP_ADDR_HI(segs[0].ds_addr));
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
bus_dmamap_sync(sc->sc_cdata.txp_rxbufs_tag,
sc->sc_cdata.txp_rxbufs_map,
BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
sc->sc_rxbufprod = RXBUF_ENTRIES - 1;
sc->sc_hostvar->hv_rx_buf_write_idx =
htole32(TXP_IDX2OFFSET(RXBUF_ENTRIES - 1));
2009-03-11 08:49:17 +00:00
return (0);
}
static void
2009-03-11 08:25:18 +00:00
txp_rxring_empty(struct txp_softc *sc)
{
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
struct txp_rx_swdesc *sd;
int cnt;
TXP_LOCK_ASSERT(sc);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if (sc->sc_rxbufs == NULL)
return;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
bus_dmamap_sync(sc->sc_cdata.txp_hostvar_tag,
sc->sc_cdata.txp_hostvar_map,
BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
/* Release allocated Rx buffers. */
cnt = 0;
while ((sd = TAILQ_FIRST(&sc->sc_busy_list)) != NULL) {
TAILQ_REMOVE(&sc->sc_busy_list, sd, sd_next);
KASSERT(sd->sd_mbuf != NULL,
("%s : Rx buffer ring corrupted", __func__));
bus_dmamap_sync(sc->sc_cdata.txp_rx_tag, sd->sd_map,
BUS_DMASYNC_POSTREAD);
bus_dmamap_unload(sc->sc_cdata.txp_rx_tag, sd->sd_map);
m_freem(sd->sd_mbuf);
sd->sd_mbuf = NULL;
TAILQ_INSERT_TAIL(&sc->sc_free_list, sd, sd_next);
cnt++;
}
}
static void
2009-03-11 08:25:18 +00:00
txp_init(void *xsc)
{
struct txp_softc *sc;
sc = xsc;
TXP_LOCK(sc);
txp_init_locked(sc);
TXP_UNLOCK(sc);
}
static void
2009-03-11 08:25:18 +00:00
txp_init_locked(struct txp_softc *sc)
{
struct ifnet *ifp;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
uint8_t *eaddr;
uint16_t p1;
uint32_t p2;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
int error;
TXP_LOCK_ASSERT(sc);
ifp = sc->sc_ifp;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if ((ifp->if_drv_flags & IFF_DRV_RUNNING) != 0)
return;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Initialize ring structure. */
txp_init_rings(sc);
/* Wakeup controller. */
WRITE_REG(sc, TXP_H2A_0, TXP_BOOTCMD_WAKEUP);
TXP_BARRIER(sc, TXP_H2A_0, 4, BUS_SPACE_BARRIER_WRITE);
/*
* It seems that earlier NV image can go back to online from
* wakeup command but newer ones require controller reset.
* So jut reset controller again.
*/
if (txp_reset(sc) != 0)
goto init_fail;
/* Download firmware. */
error = txp_download_fw(sc);
if (error != 0) {
device_printf(sc->sc_dev, "could not download firmware.\n");
goto init_fail;
}
bus_dmamap_sync(sc->sc_cdata.txp_hostvar_tag,
sc->sc_cdata.txp_hostvar_map,
BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
if ((error = txp_rxring_fill(sc)) != 0) {
device_printf(sc->sc_dev, "no memory for Rx buffers.\n");
goto init_fail;
}
bus_dmamap_sync(sc->sc_cdata.txp_hostvar_tag,
sc->sc_cdata.txp_hostvar_map,
BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
if (txp_boot(sc, STAT_WAITING_FOR_BOOT) != 0) {
device_printf(sc->sc_dev, "could not boot firmware.\n");
goto init_fail;
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/*
* Quite contrary to Typhoon T2 software functional specification,
* it seems that TXP_CMD_RECV_BUFFER_CONTROL command is not
* implemented in the firmware. This means driver should have to
* handle misaligned frames on alignment architectures. AFAIK this
* is the only controller manufactured by 3Com that has this stupid
* bug. 3Com should fix this.
*/
if (txp_command(sc, TXP_CMD_MAX_PKT_SIZE_WRITE, TXP_MAX_PKTLEN, 0, 0,
NULL, NULL, NULL, TXP_CMD_NOWAIT) != 0)
goto init_fail;
/* Undocumented command(interrupt coalescing disable?) - From Linux. */
if (txp_command(sc, TXP_CMD_FILTER_DEFINE, 0, 0, 0, NULL, NULL, NULL,
TXP_CMD_NOWAIT) != 0)
goto init_fail;
/* Set station address. */
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
eaddr = IF_LLADDR(sc->sc_ifp);
p1 = 0;
((uint8_t *)&p1)[1] = eaddr[0];
((uint8_t *)&p1)[0] = eaddr[1];
p1 = le16toh(p1);
((uint8_t *)&p2)[3] = eaddr[2];
((uint8_t *)&p2)[2] = eaddr[3];
((uint8_t *)&p2)[1] = eaddr[4];
((uint8_t *)&p2)[0] = eaddr[5];
p2 = le32toh(p2);
if (txp_command(sc, TXP_CMD_STATION_ADDRESS_WRITE, p1, p2, 0,
NULL, NULL, NULL, TXP_CMD_NOWAIT) != 0)
goto init_fail;
txp_set_filter(sc);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
txp_set_capabilities(sc);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if (txp_command(sc, TXP_CMD_CLEAR_STATISTICS, 0, 0, 0,
NULL, NULL, NULL, TXP_CMD_NOWAIT))
goto init_fail;
if (txp_command(sc, TXP_CMD_XCVR_SELECT, sc->sc_xcvr, 0, 0,
NULL, NULL, NULL, TXP_CMD_NOWAIT) != 0)
goto init_fail;
if (txp_command(sc, TXP_CMD_TX_ENABLE, 0, 0, 0, NULL, NULL, NULL,
TXP_CMD_NOWAIT) != 0)
goto init_fail;
if (txp_command(sc, TXP_CMD_RX_ENABLE, 0, 0, 0, NULL, NULL, NULL,
TXP_CMD_NOWAIT) != 0)
goto init_fail;
/* Ack all pending interrupts and enable interrupts. */
WRITE_REG(sc, TXP_ISR, TXP_INTR_ALL);
WRITE_REG(sc, TXP_IER, TXP_INTRS);
WRITE_REG(sc, TXP_IMR, TXP_INTR_NONE);
ifp->if_drv_flags |= IFF_DRV_RUNNING;
ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
callout_reset(&sc->sc_tick, hz, txp_tick, sc);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
return;
init_fail:
txp_rxring_empty(sc);
txp_init_rings(sc);
txp_reset(sc);
WRITE_REG(sc, TXP_IMR, TXP_INTR_ALL);
}
static void
2009-03-11 08:25:18 +00:00
txp_tick(void *vsc)
{
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
struct txp_softc *sc;
struct ifnet *ifp;
struct txp_rsp_desc *rsp;
struct txp_ext_desc *ext;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
int link;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
sc = vsc;
TXP_LOCK_ASSERT(sc);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
bus_dmamap_sync(sc->sc_cdata.txp_hostvar_tag,
sc->sc_cdata.txp_hostvar_map,
BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
txp_rxbuf_reclaim(sc);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
bus_dmamap_sync(sc->sc_cdata.txp_hostvar_tag,
sc->sc_cdata.txp_hostvar_map,
BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
ifp = sc->sc_ifp;
rsp = NULL;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
link = sc->sc_flags & TXP_FLAG_LINK;
if (txp_ext_command(sc, TXP_CMD_READ_STATISTICS, 0, 0, 0, NULL, 0,
&rsp, TXP_CMD_WAIT))
goto out;
if (rsp->rsp_numdesc != 6)
goto out;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
txp_stats_update(sc, rsp);
if (link == 0 && (sc->sc_flags & TXP_FLAG_LINK) != 0) {
ext = (struct txp_ext_desc *)(rsp + 1);
/* Update baudrate with resolved speed. */
if ((ext[5].ext_2 & 0x02) != 0)
ifp->if_baudrate = IF_Mbps(100);
else
ifp->if_baudrate = IF_Mbps(10);
}
out:
if (rsp != NULL)
free(rsp, M_DEVBUF);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
txp_watchdog(sc);
callout_reset(&sc->sc_tick, hz, txp_tick, sc);
}
static void
2009-03-11 08:25:18 +00:00
txp_start(struct ifnet *ifp)
{
struct txp_softc *sc;
sc = ifp->if_softc;
TXP_LOCK(sc);
txp_start_locked(ifp);
TXP_UNLOCK(sc);
}
static void
2009-03-11 08:25:18 +00:00
txp_start_locked(struct ifnet *ifp)
{
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
struct txp_softc *sc;
struct mbuf *m_head;
int enq;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
sc = ifp->if_softc;
TXP_LOCK_ASSERT(sc);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if ((ifp->if_drv_flags & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
IFF_DRV_RUNNING || (sc->sc_flags & TXP_FLAG_LINK) == 0)
return;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
for (enq = 0; !IFQ_DRV_IS_EMPTY(&ifp->if_snd); ) {
IFQ_DRV_DEQUEUE(&ifp->if_snd, m_head);
if (m_head == NULL)
break;
/*
* Pack the data into the transmit ring. If we
* don't have room, set the OACTIVE flag and wait
* for the NIC to drain the ring.
* ATM only Hi-ring is used.
*/
if (txp_encap(sc, &sc->sc_txhir, &m_head)) {
if (m_head == NULL)
break;
IFQ_DRV_PREPEND(&ifp->if_snd, m_head);
ifp->if_drv_flags |= IFF_DRV_OACTIVE;
break;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/*
* If there's a BPF listener, bounce a copy of this frame
* to him.
*/
ETHER_BPF_MTAP(ifp, m_head);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Send queued frame. */
WRITE_REG(sc, sc->sc_txhir.r_reg,
TXP_IDX2OFFSET(sc->sc_txhir.r_prod));
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if (enq > 0) {
/* Set a timeout in case the chip goes out to lunch. */
sc->sc_watchdog_timer = TXP_TX_TIMEOUT;
}
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
static int
txp_encap(struct txp_softc *sc, struct txp_tx_ring *r, struct mbuf **m_head)
{
struct txp_tx_desc *first_txd;
struct txp_frag_desc *fxd;
struct txp_swdesc *sd;
struct mbuf *m;
bus_dma_segment_t txsegs[TXP_MAXTXSEGS];
int error, i, nsegs;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
TXP_LOCK_ASSERT(sc);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
M_ASSERTPKTHDR((*m_head));
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
m = *m_head;
first_txd = r->r_desc + r->r_prod;
sd = sc->sc_txd + r->r_prod;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
error = bus_dmamap_load_mbuf_sg(sc->sc_cdata.txp_tx_tag, sd->sd_map,
*m_head, txsegs, &nsegs, 0);
if (error == EFBIG) {
m = m_collapse(*m_head, M_NOWAIT, TXP_MAXTXSEGS);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if (m == NULL) {
m_freem(*m_head);
*m_head = NULL;
return (ENOMEM);
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
*m_head = m;
error = bus_dmamap_load_mbuf_sg(sc->sc_cdata.txp_tx_tag,
sd->sd_map, *m_head, txsegs, &nsegs, 0);
if (error != 0) {
m_freem(*m_head);
*m_head = NULL;
return (error);
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
} else if (error != 0)
return (error);
if (nsegs == 0) {
m_freem(*m_head);
*m_head = NULL;
return (EIO);
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Check descriptor overrun. */
if (r->r_cnt + nsegs >= TX_ENTRIES - TXP_TXD_RESERVED) {
bus_dmamap_unload(sc->sc_cdata.txp_tx_tag, sd->sd_map);
return (ENOBUFS);
}
bus_dmamap_sync(sc->sc_cdata.txp_tx_tag, sd->sd_map,
BUS_DMASYNC_PREWRITE);
sd->sd_mbuf = m;
first_txd->tx_flags = TX_FLAGS_TYPE_DATA;
first_txd->tx_numdesc = 0;
first_txd->tx_addrlo = 0;
first_txd->tx_addrhi = 0;
first_txd->tx_totlen = 0;
first_txd->tx_pflags = 0;
r->r_cnt++;
TXP_DESC_INC(r->r_prod, TX_ENTRIES);
/* Configure Tx IP/TCP/UDP checksum offload. */
if ((m->m_pkthdr.csum_flags & CSUM_IP) != 0)
first_txd->tx_pflags |= htole32(TX_PFLAGS_IPCKSUM);
#ifdef notyet
/* XXX firmware bug. */
if ((m->m_pkthdr.csum_flags & CSUM_TCP) != 0)
first_txd->tx_pflags |= htole32(TX_PFLAGS_TCPCKSUM);
if ((m->m_pkthdr.csum_flags & CSUM_UDP) != 0)
first_txd->tx_pflags |= htole32(TX_PFLAGS_UDPCKSUM);
#endif
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Configure VLAN hardware tag insertion. */
if ((m->m_flags & M_VLANTAG) != 0)
first_txd->tx_pflags |=
htole32(TX_PFLAGS_VLAN | TX_PFLAGS_PRIO |
(bswap16(m->m_pkthdr.ether_vtag) << TX_PFLAGS_VLANTAG_S));
for (i = 0; i < nsegs; i++) {
fxd = (struct txp_frag_desc *)(r->r_desc + r->r_prod);
fxd->frag_flags = FRAG_FLAGS_TYPE_FRAG | TX_FLAGS_VALID;
fxd->frag_rsvd1 = 0;
fxd->frag_len = htole16(txsegs[i].ds_len);
fxd->frag_addrhi = htole32(TXP_ADDR_HI(txsegs[i].ds_addr));
fxd->frag_addrlo = htole32(TXP_ADDR_LO(txsegs[i].ds_addr));
fxd->frag_rsvd2 = 0;
first_txd->tx_numdesc++;
r->r_cnt++;
TXP_DESC_INC(r->r_prod, TX_ENTRIES);
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Lastly set valid flag. */
first_txd->tx_flags |= TX_FLAGS_VALID;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Sync descriptors. */
bus_dmamap_sync(r->r_tag, r->r_map,
BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
return (0);
}
/*
* Handle simple commands sent to the typhoon
*/
static int
txp_command(struct txp_softc *sc, uint16_t id, uint16_t in1, uint32_t in2,
uint32_t in3, uint16_t *out1, uint32_t *out2, uint32_t *out3, int wait)
{
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
struct txp_rsp_desc *rsp;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
rsp = NULL;
if (txp_ext_command(sc, id, in1, in2, in3, NULL, 0, &rsp, wait) != 0) {
device_printf(sc->sc_dev, "command 0x%02x failed\n", id);
return (-1);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if (wait == TXP_CMD_NOWAIT)
return (0);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
KASSERT(rsp != NULL, ("rsp is NULL!\n"));
if (out1 != NULL)
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
*out1 = le16toh(rsp->rsp_par1);
if (out2 != NULL)
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
*out2 = le32toh(rsp->rsp_par2);
if (out3 != NULL)
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
*out3 = le32toh(rsp->rsp_par3);
free(rsp, M_DEVBUF);
return (0);
}
static int
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
txp_ext_command(struct txp_softc *sc, uint16_t id, uint16_t in1, uint32_t in2,
uint32_t in3, struct txp_ext_desc *in_extp, uint8_t in_extn,
struct txp_rsp_desc **rspp, int wait)
{
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
struct txp_hostvar *hv;
struct txp_cmd_desc *cmd;
struct txp_ext_desc *ext;
uint32_t idx, i;
uint16_t seq;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
int error;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
error = 0;
hv = sc->sc_hostvar;
if (txp_cmd_desc_numfree(sc) < (in_extn + 1)) {
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
device_printf(sc->sc_dev,
"%s : out of free cmd descriptors for command 0x%02x\n",
__func__, id);
return (ENOBUFS);
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
bus_dmamap_sync(sc->sc_cdata.txp_cmdring_tag,
sc->sc_cdata.txp_cmdring_map, BUS_DMASYNC_POSTWRITE);
idx = sc->sc_cmdring.lastwrite;
cmd = (struct txp_cmd_desc *)(((uint8_t *)sc->sc_cmdring.base) + idx);
bzero(cmd, sizeof(*cmd));
cmd->cmd_numdesc = in_extn;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
seq = sc->sc_seq++;
cmd->cmd_seq = htole16(seq);
cmd->cmd_id = htole16(id);
cmd->cmd_par1 = htole16(in1);
cmd->cmd_par2 = htole32(in2);
cmd->cmd_par3 = htole32(in3);
cmd->cmd_flags = CMD_FLAGS_TYPE_CMD |
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
(wait == TXP_CMD_WAIT ? CMD_FLAGS_RESP : 0) | CMD_FLAGS_VALID;
idx += sizeof(struct txp_cmd_desc);
if (idx == sc->sc_cmdring.size)
idx = 0;
for (i = 0; i < in_extn; i++) {
ext = (struct txp_ext_desc *)(((uint8_t *)sc->sc_cmdring.base) + idx);
bcopy(in_extp, ext, sizeof(struct txp_ext_desc));
in_extp++;
idx += sizeof(struct txp_cmd_desc);
if (idx == sc->sc_cmdring.size)
idx = 0;
}
sc->sc_cmdring.lastwrite = idx;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
bus_dmamap_sync(sc->sc_cdata.txp_cmdring_tag,
sc->sc_cdata.txp_cmdring_map, BUS_DMASYNC_PREWRITE);
bus_dmamap_sync(sc->sc_cdata.txp_hostvar_tag,
sc->sc_cdata.txp_hostvar_map, BUS_DMASYNC_PREREAD |
BUS_DMASYNC_PREWRITE);
WRITE_REG(sc, TXP_H2A_2, sc->sc_cmdring.lastwrite);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
TXP_BARRIER(sc, TXP_H2A_2, 4, BUS_SPACE_BARRIER_WRITE);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if (wait == TXP_CMD_NOWAIT)
return (0);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
for (i = 0; i < TXP_TIMEOUT; i++) {
bus_dmamap_sync(sc->sc_cdata.txp_hostvar_tag,
sc->sc_cdata.txp_hostvar_map, BUS_DMASYNC_POSTREAD |
BUS_DMASYNC_POSTWRITE);
if (le32toh(hv->hv_resp_read_idx) !=
le32toh(hv->hv_resp_write_idx)) {
error = txp_response(sc, id, seq, rspp);
bus_dmamap_sync(sc->sc_cdata.txp_hostvar_tag,
sc->sc_cdata.txp_hostvar_map,
BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
if (error != 0)
return (error);
if (*rspp != NULL)
break;
}
DELAY(50);
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if (i == TXP_TIMEOUT) {
device_printf(sc->sc_dev, "command 0x%02x timedout\n", id);
error = ETIMEDOUT;
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
return (error);
}
static int
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
txp_response(struct txp_softc *sc, uint16_t id, uint16_t seq,
struct txp_rsp_desc **rspp)
{
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
struct txp_hostvar *hv;
struct txp_rsp_desc *rsp;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
uint32_t ridx;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
bus_dmamap_sync(sc->sc_cdata.txp_rspring_tag,
sc->sc_cdata.txp_rspring_map, BUS_DMASYNC_POSTREAD);
hv = sc->sc_hostvar;
ridx = le32toh(hv->hv_resp_read_idx);
while (ridx != le32toh(hv->hv_resp_write_idx)) {
rsp = (struct txp_rsp_desc *)(((uint8_t *)sc->sc_rspring.base) + ridx);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if (id == le16toh(rsp->rsp_id) &&
le16toh(rsp->rsp_seq) == seq) {
*rspp = (struct txp_rsp_desc *)malloc(
sizeof(struct txp_rsp_desc) * (rsp->rsp_numdesc + 1),
M_DEVBUF, M_NOWAIT);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if (*rspp == NULL) {
device_printf(sc->sc_dev,"%s : command 0x%02x "
"memory allocation failure\n",
__func__, id);
return (ENOMEM);
}
txp_rsp_fixup(sc, rsp, *rspp);
return (0);
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if ((rsp->rsp_flags & RSP_FLAGS_ERROR) != 0) {
device_printf(sc->sc_dev,
"%s : command 0x%02x response error!\n", __func__,
le16toh(rsp->rsp_id));
txp_rsp_fixup(sc, rsp, NULL);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
ridx = le32toh(hv->hv_resp_read_idx);
continue;
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/*
* The following unsolicited responses are handled during
* processing of TXP_CMD_READ_STATISTICS which requires
* response. Driver abuses the command to detect media
* status change.
* TXP_CMD_FILTER_DEFINE is not an unsolicited response
* but we don't process response ring in interrupt handler
* so we have to ignore this command here, otherwise
* unknown command message would be printed.
*/
switch (le16toh(rsp->rsp_id)) {
case TXP_CMD_CYCLE_STATISTICS:
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
case TXP_CMD_FILTER_DEFINE:
break;
case TXP_CMD_MEDIA_STATUS_READ:
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if ((le16toh(rsp->rsp_par1) & 0x0800) == 0) {
sc->sc_flags |= TXP_FLAG_LINK;
if_link_state_change(sc->sc_ifp,
LINK_STATE_UP);
} else {
sc->sc_flags &= ~TXP_FLAG_LINK;
if_link_state_change(sc->sc_ifp,
LINK_STATE_DOWN);
}
break;
case TXP_CMD_HELLO_RESPONSE:
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/*
* Driver should repsond to hello message but
* TXP_CMD_READ_STATISTICS is issued for every
* hz, therefore there is no need to send an
* explicit command here.
*/
device_printf(sc->sc_dev, "%s : hello\n", __func__);
break;
default:
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
device_printf(sc->sc_dev,
"%s : unknown command 0x%02x\n", __func__,
le16toh(rsp->rsp_id));
}
txp_rsp_fixup(sc, rsp, NULL);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
ridx = le32toh(hv->hv_resp_read_idx);
}
return (0);
}
static void
2009-03-11 08:25:18 +00:00
txp_rsp_fixup(struct txp_softc *sc, struct txp_rsp_desc *rsp,
struct txp_rsp_desc *dst)
{
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
struct txp_rsp_desc *src;
struct txp_hostvar *hv;
uint32_t i, ridx;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
src = rsp;
hv = sc->sc_hostvar;
ridx = le32toh(hv->hv_resp_read_idx);
for (i = 0; i < rsp->rsp_numdesc + 1; i++) {
if (dst != NULL)
bcopy(src, dst++, sizeof(struct txp_rsp_desc));
ridx += sizeof(struct txp_rsp_desc);
if (ridx == sc->sc_rspring.size) {
src = sc->sc_rspring.base;
ridx = 0;
} else
src++;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
sc->sc_rspring.lastwrite = ridx;
}
2009-03-11 08:28:24 +00:00
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
hv->hv_resp_read_idx = htole32(ridx);
}
static int
2009-03-11 08:25:18 +00:00
txp_cmd_desc_numfree(struct txp_softc *sc)
{
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
struct txp_hostvar *hv;
struct txp_boot_record *br;
uint32_t widx, ridx, nfree;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
bus_dmamap_sync(sc->sc_cdata.txp_hostvar_tag,
sc->sc_cdata.txp_hostvar_map,
BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
hv = sc->sc_hostvar;
br = sc->sc_boot;
widx = sc->sc_cmdring.lastwrite;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
ridx = le32toh(hv->hv_cmd_read_idx);
if (widx == ridx) {
/* Ring is completely free */
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
nfree = le32toh(br->br_cmd_siz) - sizeof(struct txp_cmd_desc);
} else {
if (widx > ridx)
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
nfree = le32toh(br->br_cmd_siz) -
(widx - ridx + sizeof(struct txp_cmd_desc));
else
nfree = ridx - widx - sizeof(struct txp_cmd_desc);
}
return (nfree / sizeof(struct txp_cmd_desc));
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
static int
txp_sleep(struct txp_softc *sc, int capenable)
{
uint16_t events;
int error;
events = 0;
if ((capenable & IFCAP_WOL_MAGIC) != 0)
events |= 0x01;
error = txp_command(sc, TXP_CMD_ENABLE_WAKEUP_EVENTS, events, 0, 0,
NULL, NULL, NULL, TXP_CMD_NOWAIT);
if (error == 0) {
/* Goto sleep. */
error = txp_command(sc, TXP_CMD_GOTO_SLEEP, 0, 0, 0, NULL,
NULL, NULL, TXP_CMD_NOWAIT);
if (error == 0) {
error = txp_wait(sc, STAT_SLEEPING);
if (error != 0)
device_printf(sc->sc_dev,
"unable to enter into sleep\n");
}
}
return (error);
}
static void
2009-03-11 08:25:18 +00:00
txp_stop(struct txp_softc *sc)
{
struct ifnet *ifp;
TXP_LOCK_ASSERT(sc);
ifp = sc->sc_ifp;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0)
return;
WRITE_REG(sc, TXP_IER, TXP_INTR_NONE);
WRITE_REG(sc, TXP_ISR, TXP_INTR_ALL);
ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
sc->sc_flags &= ~TXP_FLAG_LINK;
callout_stop(&sc->sc_tick);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
txp_command(sc, TXP_CMD_TX_DISABLE, 0, 0, 0, NULL, NULL, NULL,
TXP_CMD_NOWAIT);
txp_command(sc, TXP_CMD_RX_DISABLE, 0, 0, 0, NULL, NULL, NULL,
TXP_CMD_NOWAIT);
/* Save statistics for later use. */
txp_stats_save(sc);
/* Halt controller. */
txp_command(sc, TXP_CMD_HALT, 0, 0, 0, NULL, NULL, NULL,
TXP_CMD_NOWAIT);
if (txp_wait(sc, STAT_HALTED) != 0)
device_printf(sc->sc_dev, "controller halt timedout!\n");
/* Reclaim Tx/Rx buffers. */
if (sc->sc_txhir.r_cnt && (sc->sc_txhir.r_cons !=
TXP_OFFSET2IDX(le32toh(*(sc->sc_txhir.r_off)))))
txp_tx_reclaim(sc, &sc->sc_txhir);
if (sc->sc_txlor.r_cnt && (sc->sc_txlor.r_cons !=
TXP_OFFSET2IDX(le32toh(*(sc->sc_txlor.r_off)))))
txp_tx_reclaim(sc, &sc->sc_txlor);
txp_rxring_empty(sc);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
txp_init_rings(sc);
/* Reset controller and make it reload sleep image. */
txp_reset(sc);
/* Let controller boot from sleep image. */
if (txp_boot(sc, STAT_WAITING_FOR_HOST_REQUEST) != 0)
device_printf(sc->sc_dev, "could not boot sleep image\n");
txp_sleep(sc, 0);
}
static void
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
txp_watchdog(struct txp_softc *sc)
{
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
struct ifnet *ifp;
TXP_LOCK_ASSERT(sc);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if (sc->sc_watchdog_timer == 0 || --sc->sc_watchdog_timer)
return;
ifp = sc->sc_ifp;
if_printf(ifp, "watchdog timeout -- resetting\n");
ifp->if_oerrors++;
txp_stop(sc);
txp_init_locked(sc);
}
static int
2009-03-11 08:25:18 +00:00
txp_ifmedia_upd(struct ifnet *ifp)
{
struct txp_softc *sc = ifp->if_softc;
struct ifmedia *ifm = &sc->sc_ifmedia;
uint16_t new_xcvr;
TXP_LOCK(sc);
if (IFM_TYPE(ifm->ifm_media) != IFM_ETHER) {
TXP_UNLOCK(sc);
return (EINVAL);
}
if (IFM_SUBTYPE(ifm->ifm_media) == IFM_10_T) {
if ((ifm->ifm_media & IFM_GMASK) == IFM_FDX)
new_xcvr = TXP_XCVR_10_FDX;
else
new_xcvr = TXP_XCVR_10_HDX;
} else if (IFM_SUBTYPE(ifm->ifm_media) == IFM_100_TX) {
if ((ifm->ifm_media & IFM_GMASK) == IFM_FDX)
new_xcvr = TXP_XCVR_100_FDX;
else
new_xcvr = TXP_XCVR_100_HDX;
} else if (IFM_SUBTYPE(ifm->ifm_media) == IFM_AUTO) {
new_xcvr = TXP_XCVR_AUTO;
} else {
TXP_UNLOCK(sc);
return (EINVAL);
}
/* nothing to do */
if (sc->sc_xcvr == new_xcvr) {
TXP_UNLOCK(sc);
return (0);
}
txp_command(sc, TXP_CMD_XCVR_SELECT, new_xcvr, 0, 0,
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
NULL, NULL, NULL, TXP_CMD_NOWAIT);
sc->sc_xcvr = new_xcvr;
TXP_UNLOCK(sc);
return (0);
}
static void
2009-03-11 08:25:18 +00:00
txp_ifmedia_sts(struct ifnet *ifp, struct ifmediareq *ifmr)
{
struct txp_softc *sc = ifp->if_softc;
struct ifmedia *ifm = &sc->sc_ifmedia;
uint16_t bmsr, bmcr, anar, anlpar;
ifmr->ifm_status = IFM_AVALID;
ifmr->ifm_active = IFM_ETHER;
TXP_LOCK(sc);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
/* Check whether firmware is running. */
if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0)
goto bail;
if (txp_command(sc, TXP_CMD_PHY_MGMT_READ, 0, MII_BMSR, 0,
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
&bmsr, NULL, NULL, TXP_CMD_WAIT))
goto bail;
if (txp_command(sc, TXP_CMD_PHY_MGMT_READ, 0, MII_BMSR, 0,
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
&bmsr, NULL, NULL, TXP_CMD_WAIT))
goto bail;
if (txp_command(sc, TXP_CMD_PHY_MGMT_READ, 0, MII_BMCR, 0,
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
&bmcr, NULL, NULL, TXP_CMD_WAIT))
goto bail;
if (txp_command(sc, TXP_CMD_PHY_MGMT_READ, 0, MII_ANLPAR, 0,
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
&anlpar, NULL, NULL, TXP_CMD_WAIT))
goto bail;
if (txp_command(sc, TXP_CMD_PHY_MGMT_READ, 0, MII_ANAR, 0,
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
&anar, NULL, NULL, TXP_CMD_WAIT))
goto bail;
TXP_UNLOCK(sc);
if (bmsr & BMSR_LINK)
ifmr->ifm_status |= IFM_ACTIVE;
if (bmcr & BMCR_ISO) {
ifmr->ifm_active |= IFM_NONE;
ifmr->ifm_status = 0;
return;
}
if (bmcr & BMCR_LOOP)
ifmr->ifm_active |= IFM_LOOP;
if (bmcr & BMCR_AUTOEN) {
if ((bmsr & BMSR_ACOMP) == 0) {
ifmr->ifm_active |= IFM_NONE;
return;
}
anlpar &= anar;
if (anlpar & ANLPAR_TX_FD)
ifmr->ifm_active |= IFM_100_TX|IFM_FDX;
else if (anlpar & ANLPAR_T4)
ifmr->ifm_active |= IFM_100_T4;
else if (anlpar & ANLPAR_TX)
ifmr->ifm_active |= IFM_100_TX;
else if (anlpar & ANLPAR_10_FD)
ifmr->ifm_active |= IFM_10_T|IFM_FDX;
else if (anlpar & ANLPAR_10)
ifmr->ifm_active |= IFM_10_T;
else
ifmr->ifm_active |= IFM_NONE;
} else
ifmr->ifm_active = ifm->ifm_cur->ifm_media;
return;
bail:
TXP_UNLOCK(sc);
ifmr->ifm_active |= IFM_NONE;
ifmr->ifm_status &= ~IFM_AVALID;
}
#ifdef TXP_DEBUG
static void
2009-03-11 08:25:18 +00:00
txp_show_descriptor(void *d)
{
struct txp_cmd_desc *cmd = d;
struct txp_rsp_desc *rsp = d;
struct txp_tx_desc *txd = d;
struct txp_frag_desc *frgd = d;
switch (cmd->cmd_flags & CMD_FLAGS_TYPE_M) {
case CMD_FLAGS_TYPE_CMD:
/* command descriptor */
printf("[cmd flags 0x%x num %d id %d seq %d par1 0x%x par2 0x%x par3 0x%x]\n",
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
cmd->cmd_flags, cmd->cmd_numdesc, le16toh(cmd->cmd_id),
le16toh(cmd->cmd_seq), le16toh(cmd->cmd_par1),
le32toh(cmd->cmd_par2), le32toh(cmd->cmd_par3));
break;
case CMD_FLAGS_TYPE_RESP:
/* response descriptor */
printf("[rsp flags 0x%x num %d id %d seq %d par1 0x%x par2 0x%x par3 0x%x]\n",
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
rsp->rsp_flags, rsp->rsp_numdesc, le16toh(rsp->rsp_id),
le16toh(rsp->rsp_seq), le16toh(rsp->rsp_par1),
le32toh(rsp->rsp_par2), le32toh(rsp->rsp_par3));
break;
case CMD_FLAGS_TYPE_DATA:
/* data header (assuming tx for now) */
printf("[data flags 0x%x num %d totlen %d addr 0x%x/0x%x pflags 0x%x]",
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
txd->tx_flags, txd->tx_numdesc, le16toh(txd->tx_totlen),
le32toh(txd->tx_addrlo), le32toh(txd->tx_addrhi),
le32toh(txd->tx_pflags));
break;
case CMD_FLAGS_TYPE_FRAG:
/* fragment descriptor */
printf("[frag flags 0x%x rsvd1 0x%x len %d addr 0x%x/0x%x rsvd2 0x%x]",
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
frgd->frag_flags, frgd->frag_rsvd1, le16toh(frgd->frag_len),
le32toh(frgd->frag_addrlo), le32toh(frgd->frag_addrhi),
le32toh(frgd->frag_rsvd2));
break;
default:
printf("[unknown(%x) flags 0x%x num %d id %d seq %d par1 0x%x par2 0x%x par3 0x%x]\n",
cmd->cmd_flags & CMD_FLAGS_TYPE_M,
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
cmd->cmd_flags, cmd->cmd_numdesc, le16toh(cmd->cmd_id),
le16toh(cmd->cmd_seq), le16toh(cmd->cmd_par1),
le32toh(cmd->cmd_par2), le32toh(cmd->cmd_par3));
break;
}
}
#endif
static void
2009-03-11 08:25:18 +00:00
txp_set_filter(struct txp_softc *sc)
{
struct ifnet *ifp;
uint32_t crc, mchash[2];
uint16_t filter;
struct ifmultiaddr *ifma;
int mcnt;
TXP_LOCK_ASSERT(sc);
ifp = sc->sc_ifp;
filter = TXP_RXFILT_DIRECT;
if ((ifp->if_flags & IFF_BROADCAST) != 0)
filter |= TXP_RXFILT_BROADCAST;
if ((ifp->if_flags & (IFF_PROMISC | IFF_ALLMULTI)) != 0) {
if ((ifp->if_flags & IFF_ALLMULTI) != 0)
filter |= TXP_RXFILT_ALLMULTI;
if ((ifp->if_flags & IFF_PROMISC) != 0)
filter = TXP_RXFILT_PROMISC;
goto setit;
}
mchash[0] = mchash[1] = 0;
mcnt = 0;
if_maddr_rlock(ifp);
TAILQ_FOREACH(ifma, &ifp->if_multiaddrs, ifma_link) {
if (ifma->ifma_addr->sa_family != AF_LINK)
continue;
crc = ether_crc32_be(LLADDR((struct sockaddr_dl *)
ifma->ifma_addr), ETHER_ADDR_LEN);
crc &= 0x3f;
mchash[crc >> 5] |= 1 << (crc & 0x1f);
mcnt++;
}
if_maddr_runlock(ifp);
if (mcnt > 0) {
filter |= TXP_RXFILT_HASHMULTI;
txp_command(sc, TXP_CMD_MCAST_HASH_MASK_WRITE, 2, mchash[0],
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
mchash[1], NULL, NULL, NULL, TXP_CMD_NOWAIT);
}
setit:
txp_command(sc, TXP_CMD_RX_FILTER_WRITE, filter, 0, 0,
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
NULL, NULL, NULL, TXP_CMD_NOWAIT);
}
static int
txp_set_capabilities(struct txp_softc *sc)
{
struct ifnet *ifp;
uint32_t rxcap, txcap;
TXP_LOCK_ASSERT(sc);
rxcap = txcap = 0;
ifp = sc->sc_ifp;
if ((ifp->if_capenable & IFCAP_TXCSUM) != 0) {
if ((ifp->if_hwassist & CSUM_IP) != 0)
txcap |= OFFLOAD_IPCKSUM;
if ((ifp->if_hwassist & CSUM_TCP) != 0)
txcap |= OFFLOAD_TCPCKSUM;
if ((ifp->if_hwassist & CSUM_UDP) != 0)
txcap |= OFFLOAD_UDPCKSUM;
rxcap = txcap;
}
if ((ifp->if_capenable & IFCAP_RXCSUM) == 0)
rxcap &= ~(OFFLOAD_IPCKSUM | OFFLOAD_TCPCKSUM |
OFFLOAD_UDPCKSUM);
if ((ifp->if_capabilities & IFCAP_VLAN_HWTAGGING) != 0) {
rxcap |= OFFLOAD_VLAN;
txcap |= OFFLOAD_VLAN;
}
/* Tell firmware new offload configuration. */
return (txp_command(sc, TXP_CMD_OFFLOAD_WRITE, 0, txcap, rxcap, NULL,
NULL, NULL, TXP_CMD_NOWAIT));
}
static void
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
txp_stats_save(struct txp_softc *sc)
{
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
struct txp_rsp_desc *rsp;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
TXP_LOCK_ASSERT(sc);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
rsp = NULL;
if (txp_ext_command(sc, TXP_CMD_READ_STATISTICS, 0, 0, 0, NULL, 0,
&rsp, TXP_CMD_WAIT))
goto out;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if (rsp->rsp_numdesc != 6)
goto out;
txp_stats_update(sc, rsp);
out:
if (rsp != NULL)
free(rsp, M_DEVBUF);
bcopy(&sc->sc_stats, &sc->sc_ostats, sizeof(struct txp_hw_stats));
}
static void
txp_stats_update(struct txp_softc *sc, struct txp_rsp_desc *rsp)
{
struct ifnet *ifp;
struct txp_hw_stats *ostats, *stats;
struct txp_ext_desc *ext;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
TXP_LOCK_ASSERT(sc);
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
ifp = sc->sc_ifp;
ext = (struct txp_ext_desc *)(rsp + 1);
ostats = &sc->sc_ostats;
stats = &sc->sc_stats;
stats->tx_frames = ostats->tx_frames + le32toh(rsp->rsp_par2);
stats->tx_bytes = ostats->tx_bytes + (uint64_t)le32toh(rsp->rsp_par3) +
((uint64_t)le32toh(ext[0].ext_1) << 32);
stats->tx_deferred = ostats->tx_deferred + le32toh(ext[0].ext_2);
stats->tx_late_colls = ostats->tx_late_colls + le32toh(ext[0].ext_3);
stats->tx_colls = ostats->tx_colls + le32toh(ext[0].ext_4);
stats->tx_carrier_lost = ostats->tx_carrier_lost +
le32toh(ext[1].ext_1);
stats->tx_multi_colls = ostats->tx_multi_colls +
le32toh(ext[1].ext_2);
stats->tx_excess_colls = ostats->tx_excess_colls +
le32toh(ext[1].ext_3);
stats->tx_fifo_underruns = ostats->tx_fifo_underruns +
le32toh(ext[1].ext_4);
stats->tx_mcast_oflows = ostats->tx_mcast_oflows +
le32toh(ext[2].ext_1);
stats->tx_filtered = ostats->tx_filtered + le32toh(ext[2].ext_2);
stats->rx_frames = ostats->rx_frames + le32toh(ext[2].ext_3);
stats->rx_bytes = ostats->rx_bytes + (uint64_t)le32toh(ext[2].ext_4) +
((uint64_t)le32toh(ext[3].ext_1) << 32);
stats->rx_fifo_oflows = ostats->rx_fifo_oflows + le32toh(ext[3].ext_2);
stats->rx_badssd = ostats->rx_badssd + le32toh(ext[3].ext_3);
stats->rx_crcerrs = ostats->rx_crcerrs + le32toh(ext[3].ext_4);
stats->rx_lenerrs = ostats->rx_lenerrs + le32toh(ext[4].ext_1);
stats->rx_bcast_frames = ostats->rx_bcast_frames +
le32toh(ext[4].ext_2);
stats->rx_mcast_frames = ostats->rx_mcast_frames +
le32toh(ext[4].ext_3);
stats->rx_oflows = ostats->rx_oflows + le32toh(ext[4].ext_4);
stats->rx_filtered = ostats->rx_filtered + le32toh(ext[5].ext_1);
ifp->if_ierrors = stats->rx_fifo_oflows + stats->rx_badssd +
stats->rx_crcerrs + stats->rx_lenerrs + stats->rx_oflows;
ifp->if_oerrors = stats->tx_deferred + stats->tx_carrier_lost +
stats->tx_fifo_underruns + stats->tx_mcast_oflows;
ifp->if_collisions = stats->tx_late_colls + stats->tx_multi_colls +
stats->tx_excess_colls;
ifp->if_opackets = stats->tx_frames;
ifp->if_ipackets = stats->rx_frames;
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
#define TXP_SYSCTL_STAT_ADD32(c, h, n, p, d) \
SYSCTL_ADD_UINT(c, h, OID_AUTO, n, CTLFLAG_RD, p, 0, d)
#if __FreeBSD_version >= 900030
#define TXP_SYSCTL_STAT_ADD64(c, h, n, p, d) \
SYSCTL_ADD_UQUAD(c, h, OID_AUTO, n, CTLFLAG_RD, p, d)
#elif __FreeBSD_version > 800000
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
#define TXP_SYSCTL_STAT_ADD64(c, h, n, p, d) \
SYSCTL_ADD_QUAD(c, h, OID_AUTO, n, CTLFLAG_RD, p, d)
#else
#define TXP_SYSCTL_STAT_ADD64(c, h, n, p, d) \
SYSCTL_ADD_ULONG(c, h, OID_AUTO, n, CTLFLAG_RD, p, d)
#endif
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
static void
txp_sysctl_node(struct txp_softc *sc)
{
struct sysctl_ctx_list *ctx;
struct sysctl_oid_list *child, *parent;
struct sysctl_oid *tree;
struct txp_hw_stats *stats;
int error;
stats = &sc->sc_stats;
ctx = device_get_sysctl_ctx(sc->sc_dev);
child = SYSCTL_CHILDREN(device_get_sysctl_tree(sc->sc_dev));
SYSCTL_ADD_PROC(ctx, child, OID_AUTO, "process_limit",
CTLTYPE_INT | CTLFLAG_RW, &sc->sc_process_limit, 0,
sysctl_hw_txp_proc_limit, "I",
"max number of Rx events to process");
/* Pull in device tunables. */
sc->sc_process_limit = TXP_PROC_DEFAULT;
error = resource_int_value(device_get_name(sc->sc_dev),
device_get_unit(sc->sc_dev), "process_limit",
&sc->sc_process_limit);
if (error == 0) {
if (sc->sc_process_limit < TXP_PROC_MIN ||
sc->sc_process_limit > TXP_PROC_MAX) {
device_printf(sc->sc_dev,
"process_limit value out of range; "
"using default: %d\n", TXP_PROC_DEFAULT);
sc->sc_process_limit = TXP_PROC_DEFAULT;
}
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
tree = SYSCTL_ADD_NODE(ctx, child, OID_AUTO, "stats", CTLFLAG_RD,
NULL, "TXP statistics");
parent = SYSCTL_CHILDREN(tree);
/* Tx statistics. */
tree = SYSCTL_ADD_NODE(ctx, parent, OID_AUTO, "tx", CTLFLAG_RD,
NULL, "Tx MAC statistics");
child = SYSCTL_CHILDREN(tree);
TXP_SYSCTL_STAT_ADD32(ctx, child, "frames",
&stats->tx_frames, "Frames");
TXP_SYSCTL_STAT_ADD64(ctx, child, "octets",
&stats->tx_bytes, "Octets");
TXP_SYSCTL_STAT_ADD32(ctx, child, "deferred",
&stats->tx_deferred, "Deferred frames");
TXP_SYSCTL_STAT_ADD32(ctx, child, "late_colls",
&stats->tx_late_colls, "Late collisions");
TXP_SYSCTL_STAT_ADD32(ctx, child, "colls",
&stats->tx_colls, "Collisions");
TXP_SYSCTL_STAT_ADD32(ctx, child, "carrier_lost",
&stats->tx_carrier_lost, "Carrier lost");
TXP_SYSCTL_STAT_ADD32(ctx, child, "multi_colls",
&stats->tx_multi_colls, "Multiple collisions");
TXP_SYSCTL_STAT_ADD32(ctx, child, "excess_colls",
&stats->tx_excess_colls, "Excessive collisions");
TXP_SYSCTL_STAT_ADD32(ctx, child, "fifo_underruns",
&stats->tx_fifo_underruns, "FIFO underruns");
TXP_SYSCTL_STAT_ADD32(ctx, child, "mcast_oflows",
&stats->tx_mcast_oflows, "Multicast overflows");
TXP_SYSCTL_STAT_ADD32(ctx, child, "filtered",
&stats->tx_filtered, "Filtered frames");
/* Rx statistics. */
tree = SYSCTL_ADD_NODE(ctx, parent, OID_AUTO, "rx", CTLFLAG_RD,
NULL, "Rx MAC statistics");
child = SYSCTL_CHILDREN(tree);
TXP_SYSCTL_STAT_ADD32(ctx, child, "frames",
&stats->rx_frames, "Frames");
TXP_SYSCTL_STAT_ADD64(ctx, child, "octets",
&stats->rx_bytes, "Octets");
TXP_SYSCTL_STAT_ADD32(ctx, child, "fifo_oflows",
&stats->rx_fifo_oflows, "FIFO overflows");
TXP_SYSCTL_STAT_ADD32(ctx, child, "badssd",
&stats->rx_badssd, "Bad SSD");
TXP_SYSCTL_STAT_ADD32(ctx, child, "crcerrs",
&stats->rx_crcerrs, "CRC errors");
TXP_SYSCTL_STAT_ADD32(ctx, child, "lenerrs",
&stats->rx_lenerrs, "Length errors");
TXP_SYSCTL_STAT_ADD32(ctx, child, "bcast_frames",
&stats->rx_bcast_frames, "Broadcast frames");
TXP_SYSCTL_STAT_ADD32(ctx, child, "mcast_frames",
&stats->rx_mcast_frames, "Multicast frames");
TXP_SYSCTL_STAT_ADD32(ctx, child, "oflows",
&stats->rx_oflows, "Overflows");
TXP_SYSCTL_STAT_ADD32(ctx, child, "filtered",
&stats->rx_filtered, "Filtered frames");
}
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
#undef TXP_SYSCTL_STAT_ADD32
#undef TXP_SYSCTL_STAT_ADD64
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
static int
sysctl_int_range(SYSCTL_HANDLER_ARGS, int low, int high)
{
int error, value;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
if (arg1 == NULL)
return (EINVAL);
value = *(int *)arg1;
error = sysctl_handle_int(oidp, &value, 0, req);
if (error || req->newptr == NULL)
return (error);
if (value < low || value > high)
return (EINVAL);
*(int *)arg1 = value;
bus_dma(9) conversion and make txp(4) work on all architectures. o Header file cleanup. o bus_dma(9) conversion. - Removed all consumers of vtophys(9) and converted to use bus_dma(9). - Typhoon2 functional specification says the controller supports 64bit DMA addressing. However all Typhoon controllers are known to lack of DAC support so 64bit DMA support was disabled. - The hardware can't handle more 16 fragmented Tx DMA segments so teach txp(4) to collapse these segments to be less than 16. - Added Rx buffer alignment requirements(4 bytes alignment) and implemented fixup code to align receive frame. Previously txp(4) always copied Rx frame to align it on 2 byte boundary but its copy overhead is much higher than unaligned access on i386/amd64. Alignment fixup code is now applied only for strict-alignment architectures. With this change i386 and amd64 will get instant Rx performance boost. Typhoon2 datasheet mentions a command that pads arbitrary bytes in Rx buffer but that command does not work. - Nuked pointer trick in descriptor ring. This does not work on sparc64 and replaced it with bcopy. Alternatively txp(4) can embed a 32 bits index value into the descriptor and compute real buffer address but it may make code complicated. - Added endianness support code in various Tx/Rx/command/response descriptor access. With this change txp(4) should work on all architectures. o Added comments for known firmware bugs(Tx checksum offloading, TSO, VLAN stripping and Rx buffer padding control). o Prefer faster memory space register access to I/O space access. Added fall-back mechanism to use alternative I/O space access. The hardware supports both memory and I/O mapped access. Users can still force to use old I/O space access by setting hw.txp.prefer_iomap tunable to 1 in /boot/loader.conf. o Added experimental suspend/resume methods. o Nuke error prone Rx buffer handling code and implemented local buffer management with TAILQ. Be definition the controller can't pass the last received frame to host if no Rx free buffers are available to use as head and tail pointer of Rx descriptor ring can't have the same value. In that case the Rx buffer pointer in Rx buffer ring still holds a valid buffer and txp_rxbuf_reclaim() can't fill Rx buffers as the first buffer is still valid. Instead of relying on the value of Rx buffer ring, introduce local buffer management code to handle empty buffer situation. This should fix a long standing bug which completely hangs the controller under high network load. I could easily trigger the issue by sending 64 bytes UDP frames with netperf. I have no idea how this bugs was not fixed for a long time. o Converted ithread interrupt handler to filter based one. o Rearranged txp_detach routine such that it's now used for general clean-up routine. o Show sleep image version on device attach time. This will help to know what action should be taken depending on sleep image version. The version information in datasheet was wrong for newer NV images so I followed Linux which seems to correctly extract version numbers from response descriptors. o Firmware image is no longer downloaded in device attach time. Now it is reloaded whenever if_init is invoked. This is to ensure correct operation of hardware when something goes wrong. Previously the controller always run without regard to running state of firmware. This change will add additional controller initialization time but it give more robust operation as txp(4) always start off from a known state. The controller is put into sleep state until administrator explicitly up the interface. o As firmware is loaded in if_init handler, it's now possible to implement real watchdog timeout handler. When watchdog timer is expired, full-reset the controller and initialize the hardware again as most other drivers do. While I'm here use our own timer for watchdog instead of using if_watchdog/if_timer interface. o Instead of masking specific interrupts with TXP_IMR register, program TXP_IER register with the interrupts to be raised and use TXP_IMR to toggle interrupt generation. o Implemented txp_wait() to wait a specific state of a controller. o Separate boot related code from txp_download_fw() and name it txp_boot() to handle boot process. o Added bus_barrier(9) to host to ARM communication. o Added endianness to all typhoon command processing. The ARM93C always expects little-endian format of command/data. o Removed __STRICT_ALIGNMENT which is not valid on FreeBSD. __NO_STRICT_ALIGNMENT is provided for that purpose on FreeBSD. Previously __STRICT_ALIGNMENT was unconditionally defined for all architectures. o Rewrote SIOCSIFCAP ioctl handler such that each capability can be controlled by ifconfig(8). Note, disabling VLAN hardware tagging has no effect due to the bug of firmware. o Don't send TXP_CMD_CLEAR_STATISTICS to clear MAC statistics in txp_tick(). The command is not atomic. Instead, just read the statistics and reflect saved statistics to the statistics. dev.txp.%d.stats sysctl node provides detailed MAC statistics. This also reduces a lot of waste of CPU cycles as processing a command ring takes a very long time on ARM93C. Note, Rx multicast and broadcast statistics does not seem to right. It might be another bug of firmware. o Implemented link state change handling in txp_tick(). Now sending packets is allowed only after establishing a valid link. Also invoke link state change notification whenever its state is changed so pseudo drivers like lagg(4) that relies on link state can work with failover or link aggregation without hacks. if_baudrate is updated to resolved speed so SNMP agents can get correct bandwidth parameters. o Overhauled Tx routine such that it now honors number of allowable DMA segments and checks for 4 free descriptors before trying to send a frame. A frame may require 4 descriptors(1 frame descriptor, 1 or more frame descriptors, 1 TSO option descriptor, one free descriptor to prevent descriptor wrap-around) at least so it's necessary to check available free descriptors prior to setting up DMA operation. o Added a sysctl variable dev.txp.%d.process_limit to control how many received frames should be served in Rx handler. Valid ranges are 16 to 128(default 64) in unit of frames. o Added ALTQ(4) support. o Added missing IFCAP_VLAN_HWCSUM as txp(4) can offload checksum calculation as well as VLAN tag insertion/stripping. o Fixed media header length for VLAN. o Don't set if_mtu in device attach, it's already set in ether_ifattach(). o Enabled MWI. o Fixed module unload panic when bpf listeners are active. o Rearranged ethernet address programming logic such that it works on strict-alignment architectures. o Removed unused member variables in softc. o Added support for WOL. o Removed now unused TXP_PCI_LOMEM/TXP_PCI_LOIO. o Added wakeup command TXP_BOOTCMD_WAKEUP definition. o Added a new firmware version query command, TXP_CMD_READ_VERSION. o Removed volatile keyword in softc as bus_dmamap_sync(9) should take care of this. o Removed embedded union trick of a structure used to to access a pointer on LP64 systems. o Added a few TSO related definitions for struct txp_tcpseg_desc. However TSO is not used at all due to the limitation of hardware. o Redefined PKT_MAX_PKTLEN to theoretical maximum size of a frame. o Switched from bus_space_{read|write}_4 to bus_{read|write}_4. o Added a new macro TXP_DESC_INC to compute next descriptor index. Tested by: don.nasco <> gmail dot com
2009-03-12 01:14:47 +00:00
return (0);
}
static int
sysctl_hw_txp_proc_limit(SYSCTL_HANDLER_ARGS)
{
return (sysctl_int_range(oidp, arg1, arg2, req,
TXP_PROC_MIN, TXP_PROC_MAX));
}