freebsd-dev/sys/dev/gem/if_gem.c

2276 lines
62 KiB
C
Raw Normal View History

/*-
* Copyright (C) 2001 Eduardo Horvath.
* Copyright (c) 2001-2003 Thomas Moestl
* Copyright (c) 2007 Marius Strobl <marius@FreeBSD.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.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``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 THE AUTHOR 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.
*
* from: NetBSD: gem.c,v 1.21 2002/06/01 23:50:58 lukem Exp
*/
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
/*
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
* Driver for Apple GMAC, Sun ERI and Sun GEM Ethernet controllers
*/
#if 0
#define GEM_DEBUG
#endif
#if 0 /* XXX: In case of emergency, re-enable this. */
#define GEM_RINT_TIMEOUT
#endif
#include <sys/param.h>
#include <sys/systm.h>
#include <sys/bus.h>
#include <sys/callout.h>
#include <sys/endian.h>
#include <sys/mbuf.h>
#include <sys/malloc.h>
#include <sys/kernel.h>
#include <sys/lock.h>
#include <sys/module.h>
#include <sys/mutex.h>
#include <sys/socket.h>
#include <sys/sockio.h>
#include <sys/rman.h>
2002-10-26 12:32:07 +00:00
#include <net/bpf.h>
#include <net/ethernet.h>
#include <net/if.h>
#include <net/if_var.h>
#include <net/if_arp.h>
#include <net/if_dl.h>
#include <net/if_media.h>
#include <net/if_types.h>
#include <net/if_vlan_var.h>
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
#include <netinet/in.h>
#include <netinet/in_systm.h>
#include <netinet/ip.h>
#include <netinet/tcp.h>
#include <netinet/udp.h>
#include <machine/bus.h>
#include <dev/mii/mii.h>
#include <dev/mii/miivar.h>
2003-08-23 06:30:21 +00:00
#include <dev/gem/if_gemreg.h>
#include <dev/gem/if_gemvar.h>
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
CTASSERT(powerof2(GEM_NRXDESC) && GEM_NRXDESC >= 32 && GEM_NRXDESC <= 8192);
CTASSERT(powerof2(GEM_NTXDESC) && GEM_NTXDESC >= 32 && GEM_NTXDESC <= 8192);
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
#define GEM_TRIES 10000
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
/*
* The hardware supports basic TCP/UDP checksum offloading. However,
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
* the hardware doesn't compensate the checksum for UDP datagram which
* can yield to 0x0. As a safe guard, UDP checksum offload is disabled
* by default. It can be reactivated by setting special link option
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
* link0 with ifconfig(8).
*/
#define GEM_CSUM_FEATURES (CSUM_TCP)
static int gem_add_rxbuf(struct gem_softc *sc, int idx);
static int gem_bitwait(struct gem_softc *sc, u_int bank, bus_addr_t r,
uint32_t clr, uint32_t set);
static void gem_cddma_callback(void *xsc, bus_dma_segment_t *segs,
int nsegs, int error);
static int gem_disable_rx(struct gem_softc *sc);
static int gem_disable_tx(struct gem_softc *sc);
static void gem_eint(struct gem_softc *sc, u_int status);
static void gem_init(void *xsc);
static void gem_init_locked(struct gem_softc *sc);
static void gem_init_regs(struct gem_softc *sc);
static int gem_ioctl(struct ifnet *ifp, u_long cmd, caddr_t data);
static int gem_load_txmbuf(struct gem_softc *sc, struct mbuf **m_head);
static int gem_meminit(struct gem_softc *sc);
static void gem_mifinit(struct gem_softc *sc);
static void gem_reset(struct gem_softc *sc);
static int gem_reset_rx(struct gem_softc *sc);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
static void gem_reset_rxdma(struct gem_softc *sc);
static int gem_reset_tx(struct gem_softc *sc);
static u_int gem_ringsize(u_int sz);
static void gem_rint(struct gem_softc *sc);
#ifdef GEM_RINT_TIMEOUT
static void gem_rint_timeout(void *arg);
2002-07-24 01:09:38 +00:00
#endif
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
static inline void gem_rxcksum(struct mbuf *m, uint64_t flags);
static void gem_rxdrain(struct gem_softc *sc);
static void gem_setladrf(struct gem_softc *sc);
static void gem_start(struct ifnet *ifp);
static void gem_start_locked(struct ifnet *ifp);
static void gem_stop(struct ifnet *ifp, int disable);
static void gem_tick(void *arg);
static void gem_tint(struct gem_softc *sc);
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
static inline void gem_txkick(struct gem_softc *sc);
static int gem_watchdog(struct gem_softc *sc);
devclass_t gem_devclass;
DRIVER_MODULE(miibus, gem, miibus_driver, miibus_devclass, 0, 0);
MODULE_DEPEND(gem, miibus, 1, 1, 1);
#ifdef GEM_DEBUG
#include <sys/ktr.h>
#define KTR_GEM KTR_SPARE2
#endif
#define GEM_BANK1_BITWAIT(sc, r, clr, set) \
gem_bitwait((sc), GEM_RES_BANK1, (r), (clr), (set))
#define GEM_BANK2_BITWAIT(sc, r, clr, set) \
gem_bitwait((sc), GEM_RES_BANK2, (r), (clr), (set))
int
gem_attach(struct gem_softc *sc)
{
struct gem_txsoft *txs;
struct ifnet *ifp;
int error, i, phy;
uint32_t v;
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
if (bootverbose)
device_printf(sc->sc_dev, "flags=0x%x\n", sc->sc_flags);
/* Set up ifnet structure. */
ifp = sc->sc_ifp = if_alloc(IFT_ETHER);
if (ifp == NULL)
return (ENOSPC);
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
sc->sc_csum_features = GEM_CSUM_FEATURES;
ifp->if_softc = sc;
if_initname(ifp, device_get_name(sc->sc_dev),
device_get_unit(sc->sc_dev));
ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
ifp->if_start = gem_start;
ifp->if_ioctl = gem_ioctl;
ifp->if_init = gem_init;
IFQ_SET_MAXLEN(&ifp->if_snd, GEM_TXQUEUELEN);
ifp->if_snd.ifq_drv_maxlen = GEM_TXQUEUELEN;
IFQ_SET_READY(&ifp->if_snd);
callout_init_mtx(&sc->sc_tick_ch, &sc->sc_mtx, 0);
#ifdef GEM_RINT_TIMEOUT
callout_init_mtx(&sc->sc_rx_ch, &sc->sc_mtx, 0);
#endif
/* Make sure the chip is stopped. */
gem_reset(sc);
error = bus_dma_tag_create(bus_get_dma_tag(sc->sc_dev), 1, 0,
BUS_SPACE_MAXADDR_32BIT, BUS_SPACE_MAXADDR, NULL, NULL,
BUS_SPACE_MAXSIZE_32BIT, 0, BUS_SPACE_MAXSIZE_32BIT, 0, NULL,
NULL, &sc->sc_pdmatag);
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
if (error != 0)
goto fail_ifnet;
error = bus_dma_tag_create(sc->sc_pdmatag, 1, 0,
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, MCLBYTES,
1, MCLBYTES, BUS_DMA_ALLOCNOW, NULL, NULL, &sc->sc_rdmatag);
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
if (error != 0)
goto fail_ptag;
error = bus_dma_tag_create(sc->sc_pdmatag, 1, 0,
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
MCLBYTES * GEM_NTXSEGS, GEM_NTXSEGS, MCLBYTES,
BUS_DMA_ALLOCNOW, NULL, NULL, &sc->sc_tdmatag);
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
if (error != 0)
goto fail_rtag;
error = bus_dma_tag_create(sc->sc_pdmatag, PAGE_SIZE, 0,
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
sizeof(struct gem_control_data), 1,
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
sizeof(struct gem_control_data), 0,
NULL, NULL, &sc->sc_cdmatag);
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
if (error != 0)
goto fail_ttag;
/*
* Allocate the control data structures, create and load the
* DMA map for it.
*/
if ((error = bus_dmamem_alloc(sc->sc_cdmatag,
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
(void **)&sc->sc_control_data,
BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO,
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
&sc->sc_cddmamap)) != 0) {
device_printf(sc->sc_dev,
"unable to allocate control data, error = %d\n", error);
goto fail_ctag;
}
sc->sc_cddma = 0;
if ((error = bus_dmamap_load(sc->sc_cdmatag, sc->sc_cddmamap,
sc->sc_control_data, sizeof(struct gem_control_data),
gem_cddma_callback, sc, 0)) != 0 || sc->sc_cddma == 0) {
device_printf(sc->sc_dev,
"unable to load control data DMA map, error = %d\n",
error);
goto fail_cmem;
}
/*
* Initialize the transmit job descriptors.
*/
STAILQ_INIT(&sc->sc_txfreeq);
STAILQ_INIT(&sc->sc_txdirtyq);
/*
* Create the transmit buffer DMA maps.
*/
error = ENOMEM;
for (i = 0; i < GEM_TXQUEUELEN; i++) {
txs = &sc->sc_txsoft[i];
txs->txs_mbuf = NULL;
txs->txs_ndescs = 0;
if ((error = bus_dmamap_create(sc->sc_tdmatag, 0,
&txs->txs_dmamap)) != 0) {
device_printf(sc->sc_dev,
"unable to create TX DMA map %d, error = %d\n",
i, error);
goto fail_txd;
}
STAILQ_INSERT_TAIL(&sc->sc_txfreeq, txs, txs_q);
}
/*
* Create the receive buffer DMA maps.
*/
for (i = 0; i < GEM_NRXDESC; i++) {
if ((error = bus_dmamap_create(sc->sc_rdmatag, 0,
&sc->sc_rxsoft[i].rxs_dmamap)) != 0) {
device_printf(sc->sc_dev,
"unable to create RX DMA map %d, error = %d\n",
i, error);
goto fail_rxd;
}
sc->sc_rxsoft[i].rxs_mbuf = NULL;
}
/* Bypass probing PHYs if we already know for sure to use a SERDES. */
if ((sc->sc_flags & GEM_SERDES) != 0)
goto serdes;
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
/* Bad things will happen when touching this register on ERI. */
if (sc->sc_variant != GEM_SUN_ERI) {
GEM_BANK1_WRITE_4(sc, GEM_MII_DATAPATH_MODE,
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
GEM_MII_DATAPATH_MII);
GEM_BANK1_BARRIER(sc, GEM_MII_DATAPATH_MODE, 4,
BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE);
}
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
gem_mifinit(sc);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
/*
* Look for an external PHY.
*/
error = ENXIO;
v = GEM_BANK1_READ_4(sc, GEM_MIF_CONFIG);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
if ((v & GEM_MIF_CONFIG_MDI1) != 0) {
v |= GEM_MIF_CONFIG_PHY_SEL;
GEM_BANK1_WRITE_4(sc, GEM_MIF_CONFIG, v);
GEM_BANK1_BARRIER(sc, GEM_MIF_CONFIG, 4,
BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
switch (sc->sc_variant) {
case GEM_SUN_ERI:
phy = GEM_PHYAD_EXTERNAL;
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
break;
default:
phy = MII_PHY_ANY;
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
break;
}
error = mii_attach(sc->sc_dev, &sc->sc_miibus, ifp,
gem_mediachange, gem_mediastatus, BMSR_DEFCAPMASK, phy,
MII_OFFSET_ANY, MIIF_DOPAUSE);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
}
/*
* Fall back on an internal PHY if no external PHY was found.
* Note that with Apple (K2) GMACs GEM_MIF_CONFIG_MDI0 can't be
* trusted when the firmware has powered down the chip.
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
*/
if (error != 0 &&
((v & GEM_MIF_CONFIG_MDI0) != 0 || GEM_IS_APPLE(sc))) {
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
v &= ~GEM_MIF_CONFIG_PHY_SEL;
GEM_BANK1_WRITE_4(sc, GEM_MIF_CONFIG, v);
GEM_BANK1_BARRIER(sc, GEM_MIF_CONFIG, 4,
BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
switch (sc->sc_variant) {
case GEM_SUN_ERI:
case GEM_APPLE_K2_GMAC:
phy = GEM_PHYAD_INTERNAL;
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
break;
case GEM_APPLE_GMAC:
phy = GEM_PHYAD_EXTERNAL;
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
break;
default:
phy = MII_PHY_ANY;
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
break;
}
error = mii_attach(sc->sc_dev, &sc->sc_miibus, ifp,
gem_mediachange, gem_mediastatus, BMSR_DEFCAPMASK, phy,
MII_OFFSET_ANY, MIIF_DOPAUSE);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
}
/*
* Try the external PCS SERDES if we didn't find any PHYs.
*/
if (error != 0 && sc->sc_variant == GEM_SUN_GEM) {
serdes:
GEM_BANK1_WRITE_4(sc, GEM_MII_DATAPATH_MODE,
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
GEM_MII_DATAPATH_SERDES);
GEM_BANK1_BARRIER(sc, GEM_MII_DATAPATH_MODE, 4,
BUS_SPACE_BARRIER_WRITE);
GEM_BANK1_WRITE_4(sc, GEM_MII_SLINK_CONTROL,
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
GEM_MII_SLINK_LOOPBACK | GEM_MII_SLINK_EN_SYNC_D);
GEM_BANK1_BARRIER(sc, GEM_MII_SLINK_CONTROL, 4,
BUS_SPACE_BARRIER_WRITE);
GEM_BANK1_WRITE_4(sc, GEM_MII_CONFIG, GEM_MII_CONFIG_ENABLE);
GEM_BANK1_BARRIER(sc, GEM_MII_CONFIG, 4,
BUS_SPACE_BARRIER_WRITE);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
sc->sc_flags |= GEM_SERDES;
error = mii_attach(sc->sc_dev, &sc->sc_miibus, ifp,
gem_mediachange, gem_mediastatus, BMSR_DEFCAPMASK,
GEM_PHYAD_EXTERNAL, MII_OFFSET_ANY, MIIF_DOPAUSE);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
}
if (error != 0) {
device_printf(sc->sc_dev, "attaching PHYs failed\n");
goto fail_rxd;
}
sc->sc_mii = device_get_softc(sc->sc_miibus);
/*
* From this point forward, the attachment cannot fail. A failure
* before this point releases all resources that may have been
* allocated.
*/
/* Get RX FIFO size. */
sc->sc_rxfifosize = 64 *
GEM_BANK1_READ_4(sc, GEM_RX_FIFO_SIZE);
/* Get TX FIFO size. */
v = GEM_BANK1_READ_4(sc, GEM_TX_FIFO_SIZE);
device_printf(sc->sc_dev, "%ukB RX FIFO, %ukB TX FIFO\n",
sc->sc_rxfifosize / 1024, v / 16);
/* Attach the interface. */
ether_ifattach(ifp, sc->sc_enaddr);
/*
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
* Tell the upper layer(s) we support long frames/checksum offloads.
*/
ifp->if_hdrlen = sizeof(struct ether_vlan_header);
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
ifp->if_capabilities |= IFCAP_VLAN_MTU | IFCAP_HWCSUM;
ifp->if_hwassist |= sc->sc_csum_features;
ifp->if_capenable |= IFCAP_VLAN_MTU | IFCAP_HWCSUM;
return (0);
/*
* Free any resources we've allocated during the failed attach
* attempt. Do this in reverse order and fall through.
*/
fail_rxd:
for (i = 0; i < GEM_NRXDESC; i++)
if (sc->sc_rxsoft[i].rxs_dmamap != NULL)
bus_dmamap_destroy(sc->sc_rdmatag,
sc->sc_rxsoft[i].rxs_dmamap);
fail_txd:
for (i = 0; i < GEM_TXQUEUELEN; i++)
if (sc->sc_txsoft[i].txs_dmamap != NULL)
bus_dmamap_destroy(sc->sc_tdmatag,
sc->sc_txsoft[i].txs_dmamap);
bus_dmamap_unload(sc->sc_cdmatag, sc->sc_cddmamap);
fail_cmem:
bus_dmamem_free(sc->sc_cdmatag, sc->sc_control_data,
sc->sc_cddmamap);
fail_ctag:
bus_dma_tag_destroy(sc->sc_cdmatag);
fail_ttag:
bus_dma_tag_destroy(sc->sc_tdmatag);
fail_rtag:
bus_dma_tag_destroy(sc->sc_rdmatag);
fail_ptag:
bus_dma_tag_destroy(sc->sc_pdmatag);
fail_ifnet:
if_free(ifp);
return (error);
}
void
gem_detach(struct gem_softc *sc)
{
struct ifnet *ifp = sc->sc_ifp;
int i;
ether_ifdetach(ifp);
GEM_LOCK(sc);
gem_stop(ifp, 1);
GEM_UNLOCK(sc);
callout_drain(&sc->sc_tick_ch);
#ifdef GEM_RINT_TIMEOUT
callout_drain(&sc->sc_rx_ch);
#endif
if_free(ifp);
device_delete_child(sc->sc_dev, sc->sc_miibus);
for (i = 0; i < GEM_NRXDESC; i++)
if (sc->sc_rxsoft[i].rxs_dmamap != NULL)
bus_dmamap_destroy(sc->sc_rdmatag,
sc->sc_rxsoft[i].rxs_dmamap);
for (i = 0; i < GEM_TXQUEUELEN; i++)
if (sc->sc_txsoft[i].txs_dmamap != NULL)
bus_dmamap_destroy(sc->sc_tdmatag,
sc->sc_txsoft[i].txs_dmamap);
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
GEM_CDSYNC(sc, BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
bus_dmamap_unload(sc->sc_cdmatag, sc->sc_cddmamap);
bus_dmamem_free(sc->sc_cdmatag, sc->sc_control_data,
sc->sc_cddmamap);
bus_dma_tag_destroy(sc->sc_cdmatag);
bus_dma_tag_destroy(sc->sc_tdmatag);
bus_dma_tag_destroy(sc->sc_rdmatag);
bus_dma_tag_destroy(sc->sc_pdmatag);
}
void
gem_suspend(struct gem_softc *sc)
{
struct ifnet *ifp = sc->sc_ifp;
GEM_LOCK(sc);
gem_stop(ifp, 0);
GEM_UNLOCK(sc);
}
void
gem_resume(struct gem_softc *sc)
{
struct ifnet *ifp = sc->sc_ifp;
GEM_LOCK(sc);
/*
* On resume all registers have to be initialized again like
* after power-on.
*/
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
sc->sc_flags &= ~GEM_INITED;
if (ifp->if_flags & IFF_UP)
gem_init_locked(sc);
GEM_UNLOCK(sc);
}
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
static inline void
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
gem_rxcksum(struct mbuf *m, uint64_t flags)
{
struct ether_header *eh;
struct ip *ip;
struct udphdr *uh;
uint16_t *opts;
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
int32_t hlen, len, pktlen;
uint32_t temp32;
uint16_t cksum;
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
pktlen = m->m_pkthdr.len;
if (pktlen < sizeof(struct ether_header) + sizeof(struct ip))
return;
eh = mtod(m, struct ether_header *);
if (eh->ether_type != htons(ETHERTYPE_IP))
return;
ip = (struct ip *)(eh + 1);
if (ip->ip_v != IPVERSION)
return;
hlen = ip->ip_hl << 2;
pktlen -= sizeof(struct ether_header);
if (hlen < sizeof(struct ip))
return;
if (ntohs(ip->ip_len) < hlen)
return;
if (ntohs(ip->ip_len) != pktlen)
return;
if (ip->ip_off & htons(IP_MF | IP_OFFMASK))
return; /* Cannot handle fragmented packet. */
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
switch (ip->ip_p) {
case IPPROTO_TCP:
if (pktlen < (hlen + sizeof(struct tcphdr)))
return;
break;
case IPPROTO_UDP:
if (pktlen < (hlen + sizeof(struct udphdr)))
return;
uh = (struct udphdr *)((uint8_t *)ip + hlen);
if (uh->uh_sum == 0)
return; /* no checksum */
break;
default:
return;
}
cksum = ~(flags & GEM_RD_CHECKSUM);
/* checksum fixup for IP options */
len = hlen - sizeof(struct ip);
if (len > 0) {
opts = (uint16_t *)(ip + 1);
for (; len > 0; len -= sizeof(uint16_t), opts++) {
temp32 = cksum - *opts;
temp32 = (temp32 >> 16) + (temp32 & 65535);
cksum = temp32 & 65535;
}
}
m->m_pkthdr.csum_flags |= CSUM_DATA_VALID;
m->m_pkthdr.csum_data = cksum;
}
static void
gem_cddma_callback(void *xsc, bus_dma_segment_t *segs, int nsegs, int error)
{
struct gem_softc *sc = xsc;
if (error != 0)
return;
if (nsegs != 1)
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
panic("%s: bad control buffer segment count", __func__);
sc->sc_cddma = segs[0].ds_addr;
}
static void
gem_tick(void *arg)
{
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
struct gem_softc *sc = arg;
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
struct ifnet *ifp = sc->sc_ifp;
uint32_t v;
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
GEM_LOCK_ASSERT(sc, MA_OWNED);
/*
* Unload collision and error counters.
*/
if_inc_counter(ifp, IFCOUNTER_COLLISIONS,
GEM_BANK1_READ_4(sc, GEM_MAC_NORM_COLL_CNT) +
GEM_BANK1_READ_4(sc, GEM_MAC_FIRST_COLL_CNT));
v = GEM_BANK1_READ_4(sc, GEM_MAC_EXCESS_COLL_CNT) +
GEM_BANK1_READ_4(sc, GEM_MAC_LATE_COLL_CNT);
if_inc_counter(ifp, IFCOUNTER_COLLISIONS, v);
if_inc_counter(ifp, IFCOUNTER_OERRORS, v);
if_inc_counter(ifp, IFCOUNTER_IERRORS,
GEM_BANK1_READ_4(sc, GEM_MAC_RX_LEN_ERR_CNT) +
GEM_BANK1_READ_4(sc, GEM_MAC_RX_ALIGN_ERR) +
GEM_BANK1_READ_4(sc, GEM_MAC_RX_CRC_ERR_CNT) +
GEM_BANK1_READ_4(sc, GEM_MAC_RX_CODE_VIOL));
/*
* Then clear the hardware counters.
*/
GEM_BANK1_WRITE_4(sc, GEM_MAC_NORM_COLL_CNT, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_FIRST_COLL_CNT, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_EXCESS_COLL_CNT, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_LATE_COLL_CNT, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_RX_LEN_ERR_CNT, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_RX_ALIGN_ERR, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_RX_CRC_ERR_CNT, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_RX_CODE_VIOL, 0);
mii_tick(sc->sc_mii);
if (gem_watchdog(sc) == EJUSTRETURN)
return;
callout_reset(&sc->sc_tick_ch, hz, gem_tick, sc);
}
static int
gem_bitwait(struct gem_softc *sc, u_int bank, bus_addr_t r, uint32_t clr,
uint32_t set)
{
int i;
uint32_t reg;
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
for (i = GEM_TRIES; i--; DELAY(100)) {
reg = GEM_BANKN_READ_M(bank, 4, sc, r);
if ((reg & clr) == 0 && (reg & set) == set)
return (1);
}
return (0);
}
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
static void
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
gem_reset(struct gem_softc *sc)
{
#ifdef GEM_DEBUG
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
CTR2(KTR_GEM, "%s: %s", device_get_name(sc->sc_dev), __func__);
#endif
gem_reset_rx(sc);
gem_reset_tx(sc);
/* Do a full reset. */
GEM_BANK2_WRITE_4(sc, GEM_RESET, GEM_RESET_RX | GEM_RESET_TX |
(sc->sc_variant == GEM_SUN_ERI ? GEM_ERI_CACHE_LINE_SIZE <<
GEM_RESET_CLSZ_SHFT : 0));
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
GEM_BANK2_BARRIER(sc, GEM_RESET, 4,
BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE);
if (!GEM_BANK2_BITWAIT(sc, GEM_RESET, GEM_RESET_RX | GEM_RESET_TX, 0))
device_printf(sc->sc_dev, "cannot reset device\n");
}
static void
gem_rxdrain(struct gem_softc *sc)
{
struct gem_rxsoft *rxs;
int i;
for (i = 0; i < GEM_NRXDESC; i++) {
rxs = &sc->sc_rxsoft[i];
if (rxs->rxs_mbuf != NULL) {
bus_dmamap_sync(sc->sc_rdmatag, rxs->rxs_dmamap,
BUS_DMASYNC_POSTREAD);
bus_dmamap_unload(sc->sc_rdmatag, rxs->rxs_dmamap);
m_freem(rxs->rxs_mbuf);
rxs->rxs_mbuf = NULL;
}
}
}
static void
gem_stop(struct ifnet *ifp, int disable)
{
struct gem_softc *sc = ifp->if_softc;
struct gem_txsoft *txs;
#ifdef GEM_DEBUG
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
CTR2(KTR_GEM, "%s: %s", device_get_name(sc->sc_dev), __func__);
#endif
callout_stop(&sc->sc_tick_ch);
#ifdef GEM_RINT_TIMEOUT
callout_stop(&sc->sc_rx_ch);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
#endif
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
gem_reset_tx(sc);
gem_reset_rx(sc);
/*
* Release any queued transmit buffers.
*/
while ((txs = STAILQ_FIRST(&sc->sc_txdirtyq)) != NULL) {
STAILQ_REMOVE_HEAD(&sc->sc_txdirtyq, txs_q);
if (txs->txs_ndescs != 0) {
bus_dmamap_sync(sc->sc_tdmatag, txs->txs_dmamap,
BUS_DMASYNC_POSTWRITE);
bus_dmamap_unload(sc->sc_tdmatag, txs->txs_dmamap);
if (txs->txs_mbuf != NULL) {
m_freem(txs->txs_mbuf);
txs->txs_mbuf = NULL;
}
}
STAILQ_INSERT_TAIL(&sc->sc_txfreeq, txs, txs_q);
}
if (disable)
gem_rxdrain(sc);
/*
* Mark the interface down and cancel the watchdog timer.
*/
ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
sc->sc_flags &= ~GEM_LINK;
sc->sc_wdog_timer = 0;
}
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
static int
gem_reset_rx(struct gem_softc *sc)
{
/*
* Resetting while DMA is in progress can cause a bus hang, so we
* disable DMA first.
*/
(void)gem_disable_rx(sc);
GEM_BANK1_WRITE_4(sc, GEM_RX_CONFIG, 0);
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
GEM_BANK1_BARRIER(sc, GEM_RX_CONFIG, 4,
BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE);
if (!GEM_BANK1_BITWAIT(sc, GEM_RX_CONFIG, GEM_RX_CONFIG_RXDMA_EN, 0))
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
device_printf(sc->sc_dev, "cannot disable RX DMA\n");
/* Wait 5ms extra. */
DELAY(5000);
/* Reset the ERX. */
GEM_BANK2_WRITE_4(sc, GEM_RESET, GEM_RESET_RX |
(sc->sc_variant == GEM_SUN_ERI ? GEM_ERI_CACHE_LINE_SIZE <<
GEM_RESET_CLSZ_SHFT : 0));
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
GEM_BANK2_BARRIER(sc, GEM_RESET, 4,
BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE);
if (!GEM_BANK2_BITWAIT(sc, GEM_RESET, GEM_RESET_RX, 0)) {
device_printf(sc->sc_dev, "cannot reset receiver\n");
return (1);
}
/* Finally, reset RX MAC. */
GEM_BANK1_WRITE_4(sc, GEM_MAC_RXRESET, 1);
GEM_BANK1_BARRIER(sc, GEM_MAC_RXRESET, 4,
BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE);
if (!GEM_BANK1_BITWAIT(sc, GEM_MAC_RXRESET, 1, 0)) {
device_printf(sc->sc_dev, "cannot reset RX MAC\n");
return (1);
}
return (0);
}
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
/*
* Reset the receiver DMA engine.
*
* Intended to be used in case of GEM_INTR_RX_TAG_ERR, GEM_MAC_RX_OVERFLOW
* etc in order to reset the receiver DMA engine only and not do a full
* reset which amongst others also downs the link and clears the FIFOs.
*/
static void
gem_reset_rxdma(struct gem_softc *sc)
{
int i;
if (gem_reset_rx(sc) != 0) {
sc->sc_ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
return (gem_init_locked(sc));
}
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
for (i = 0; i < GEM_NRXDESC; i++)
if (sc->sc_rxsoft[i].rxs_mbuf != NULL)
GEM_UPDATE_RXDESC(sc, i);
sc->sc_rxptr = 0;
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
GEM_CDSYNC(sc, BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
/* NOTE: we use only 32-bit DMA addresses here. */
GEM_BANK1_WRITE_4(sc, GEM_RX_RING_PTR_HI, 0);
GEM_BANK1_WRITE_4(sc, GEM_RX_RING_PTR_LO, GEM_CDRXADDR(sc, 0));
GEM_BANK1_WRITE_4(sc, GEM_RX_KICK, GEM_NRXDESC - 4);
GEM_BANK1_WRITE_4(sc, GEM_RX_CONFIG,
gem_ringsize(GEM_NRXDESC /* XXX */) |
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
((ETHER_HDR_LEN + sizeof(struct ip)) <<
GEM_RX_CONFIG_CXM_START_SHFT) |
(GEM_THRSH_1024 << GEM_RX_CONFIG_FIFO_THRS_SHIFT) |
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
(ETHER_ALIGN << GEM_RX_CONFIG_FBOFF_SHFT));
/* Adjusting for the SBus clock probably isn't worth the fuzz. */
GEM_BANK1_WRITE_4(sc, GEM_RX_BLANKING,
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
((6 * (sc->sc_flags & GEM_PCI66) != 0 ? 2 : 1) <<
GEM_RX_BLANKING_TIME_SHIFT) | 6);
GEM_BANK1_WRITE_4(sc, GEM_RX_PAUSE_THRESH,
(3 * sc->sc_rxfifosize / 256) |
((sc->sc_rxfifosize / 256) << 12));
GEM_BANK1_WRITE_4(sc, GEM_RX_CONFIG,
GEM_BANK1_READ_4(sc, GEM_RX_CONFIG) | GEM_RX_CONFIG_RXDMA_EN);
GEM_BANK1_WRITE_4(sc, GEM_MAC_RX_MASK,
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
GEM_MAC_RX_DONE | GEM_MAC_RX_FRAME_CNT);
/*
* Clear the RX filter and reprogram it. This will also set the
* current RX MAC configuration and enable it.
*/
gem_setladrf(sc);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
}
static int
gem_reset_tx(struct gem_softc *sc)
{
/*
* Resetting while DMA is in progress can cause a bus hang, so we
* disable DMA first.
*/
(void)gem_disable_tx(sc);
GEM_BANK1_WRITE_4(sc, GEM_TX_CONFIG, 0);
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
GEM_BANK1_BARRIER(sc, GEM_TX_CONFIG, 4,
BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE);
if (!GEM_BANK1_BITWAIT(sc, GEM_TX_CONFIG, GEM_TX_CONFIG_TXDMA_EN, 0))
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
device_printf(sc->sc_dev, "cannot disable TX DMA\n");
/* Wait 5ms extra. */
DELAY(5000);
/* Finally, reset the ETX. */
GEM_BANK2_WRITE_4(sc, GEM_RESET, GEM_RESET_TX |
(sc->sc_variant == GEM_SUN_ERI ? GEM_ERI_CACHE_LINE_SIZE <<
GEM_RESET_CLSZ_SHFT : 0));
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
GEM_BANK2_BARRIER(sc, GEM_RESET, 4,
BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE);
if (!GEM_BANK2_BITWAIT(sc, GEM_RESET, GEM_RESET_TX, 0)) {
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
device_printf(sc->sc_dev, "cannot reset transmitter\n");
return (1);
}
return (0);
}
static int
gem_disable_rx(struct gem_softc *sc)
{
GEM_BANK1_WRITE_4(sc, GEM_MAC_RX_CONFIG,
GEM_BANK1_READ_4(sc, GEM_MAC_RX_CONFIG) & ~GEM_MAC_RX_ENABLE);
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
GEM_BANK1_BARRIER(sc, GEM_MAC_RX_CONFIG, 4,
BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE);
if (GEM_BANK1_BITWAIT(sc, GEM_MAC_RX_CONFIG, GEM_MAC_RX_ENABLE, 0))
return (1);
device_printf(sc->sc_dev, "cannot disable RX MAC\n");
return (0);
}
static int
gem_disable_tx(struct gem_softc *sc)
{
GEM_BANK1_WRITE_4(sc, GEM_MAC_TX_CONFIG,
GEM_BANK1_READ_4(sc, GEM_MAC_TX_CONFIG) & ~GEM_MAC_TX_ENABLE);
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
GEM_BANK1_BARRIER(sc, GEM_MAC_TX_CONFIG, 4,
BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE);
if (GEM_BANK1_BITWAIT(sc, GEM_MAC_TX_CONFIG, GEM_MAC_TX_ENABLE, 0))
return (1);
device_printf(sc->sc_dev, "cannot disable TX MAC\n");
return (0);
}
static int
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
gem_meminit(struct gem_softc *sc)
{
struct gem_rxsoft *rxs;
int error, i;
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
GEM_LOCK_ASSERT(sc, MA_OWNED);
/*
* Initialize the transmit descriptor ring.
*/
for (i = 0; i < GEM_NTXDESC; i++) {
sc->sc_txdescs[i].gd_flags = 0;
sc->sc_txdescs[i].gd_addr = 0;
}
sc->sc_txfree = GEM_MAXTXFREE;
sc->sc_txnext = 0;
sc->sc_txwin = 0;
/*
* Initialize the receive descriptor and receive job
* descriptor rings.
*/
for (i = 0; i < GEM_NRXDESC; i++) {
rxs = &sc->sc_rxsoft[i];
if (rxs->rxs_mbuf == NULL) {
if ((error = gem_add_rxbuf(sc, i)) != 0) {
device_printf(sc->sc_dev,
"unable to allocate or map RX buffer %d, "
"error = %d\n", i, error);
/*
* XXX we should attempt to run with fewer
* receive buffers instead of just failing.
*/
gem_rxdrain(sc);
return (1);
}
} else
GEM_INIT_RXDESC(sc, i);
}
sc->sc_rxptr = 0;
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
GEM_CDSYNC(sc, BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
return (0);
}
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
static u_int
gem_ringsize(u_int sz)
{
switch (sz) {
case 32:
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
return (GEM_RING_SZ_32);
case 64:
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
return (GEM_RING_SZ_64);
case 128:
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
return (GEM_RING_SZ_128);
case 256:
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
return (GEM_RING_SZ_256);
case 512:
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
return (GEM_RING_SZ_512);
case 1024:
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
return (GEM_RING_SZ_1024);
case 2048:
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
return (GEM_RING_SZ_2048);
case 4096:
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
return (GEM_RING_SZ_4096);
case 8192:
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
return (GEM_RING_SZ_8192);
default:
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
printf("%s: invalid ring size %d\n", __func__, sz);
return (GEM_RING_SZ_32);
}
}
static void
gem_init(void *xsc)
{
struct gem_softc *sc = xsc;
GEM_LOCK(sc);
gem_init_locked(sc);
GEM_UNLOCK(sc);
}
/*
* Initialization of interface; set up initialization block
* and transmit/receive descriptor rings.
*/
static void
gem_init_locked(struct gem_softc *sc)
{
struct ifnet *ifp = sc->sc_ifp;
uint32_t v;
GEM_LOCK_ASSERT(sc, MA_OWNED);
if ((ifp->if_drv_flags & IFF_DRV_RUNNING) != 0)
return;
#ifdef GEM_DEBUG
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
CTR2(KTR_GEM, "%s: %s: calling stop", device_get_name(sc->sc_dev),
__func__);
#endif
/*
* Initialization sequence. The numbered steps below correspond
* to the sequence outlined in section 6.3.5.1 in the Ethernet
* Channel Engine manual (part of the PCIO manual).
* See also the STP2002-STQ document from Sun Microsystems.
*/
/* step 1 & 2. Reset the Ethernet Channel. */
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
gem_stop(ifp, 0);
gem_reset(sc);
#ifdef GEM_DEBUG
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
CTR2(KTR_GEM, "%s: %s: restarting", device_get_name(sc->sc_dev),
__func__);
#endif
if ((sc->sc_flags & GEM_SERDES) == 0)
/* Re-initialize the MIF. */
gem_mifinit(sc);
/* step 3. Setup data structures in host memory. */
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
if (gem_meminit(sc) != 0)
return;
/* step 4. TX MAC registers & counters */
gem_init_regs(sc);
/* step 5. RX MAC registers & counters */
/* step 6 & 7. Program Descriptor Ring Base Addresses. */
/* NOTE: we use only 32-bit DMA addresses here. */
GEM_BANK1_WRITE_4(sc, GEM_TX_RING_PTR_HI, 0);
GEM_BANK1_WRITE_4(sc, GEM_TX_RING_PTR_LO, GEM_CDTXADDR(sc, 0));
GEM_BANK1_WRITE_4(sc, GEM_RX_RING_PTR_HI, 0);
GEM_BANK1_WRITE_4(sc, GEM_RX_RING_PTR_LO, GEM_CDRXADDR(sc, 0));
#ifdef GEM_DEBUG
CTR3(KTR_GEM, "loading RX ring %lx, TX ring %lx, cddma %lx",
GEM_CDRXADDR(sc, 0), GEM_CDTXADDR(sc, 0), sc->sc_cddma);
#endif
/* step 8. Global Configuration & Interrupt Mask */
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
/*
* Set the internal arbitration to "infinite" bursts of the
* maximum length of 31 * 64 bytes so DMA transfers aren't
* split up in cache line size chunks. This greatly improves
* RX performance.
* Enable silicon bug workarounds for the Apple variants.
*/
GEM_BANK1_WRITE_4(sc, GEM_CONFIG,
GEM_CONFIG_TXDMA_LIMIT | GEM_CONFIG_RXDMA_LIMIT |
((sc->sc_flags & GEM_PCI) != 0 ? GEM_CONFIG_BURST_INF :
GEM_CONFIG_BURST_64) | (GEM_IS_APPLE(sc) ?
GEM_CONFIG_RONPAULBIT | GEM_CONFIG_BUG2FIX : 0));
GEM_BANK1_WRITE_4(sc, GEM_INTMASK,
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
~(GEM_INTR_TX_INTME | GEM_INTR_TX_EMPTY | GEM_INTR_RX_DONE |
GEM_INTR_RX_NOBUF | GEM_INTR_RX_TAG_ERR | GEM_INTR_PERR |
GEM_INTR_BERR
#ifdef GEM_DEBUG
| GEM_INTR_PCS | GEM_INTR_MIF
#endif
));
GEM_BANK1_WRITE_4(sc, GEM_MAC_RX_MASK,
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
GEM_MAC_RX_DONE | GEM_MAC_RX_FRAME_CNT);
GEM_BANK1_WRITE_4(sc, GEM_MAC_TX_MASK,
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
GEM_MAC_TX_XMIT_DONE | GEM_MAC_TX_DEFER_EXP |
GEM_MAC_TX_PEAK_EXP);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
#ifdef GEM_DEBUG
GEM_BANK1_WRITE_4(sc, GEM_MAC_CONTROL_MASK,
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
~(GEM_MAC_PAUSED | GEM_MAC_PAUSE | GEM_MAC_RESUME));
#else
GEM_BANK1_WRITE_4(sc, GEM_MAC_CONTROL_MASK,
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
GEM_MAC_PAUSED | GEM_MAC_PAUSE | GEM_MAC_RESUME);
#endif
/* step 9. ETX Configuration: use mostly default values. */
/* Enable DMA. */
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
v = gem_ringsize(GEM_NTXDESC);
/* Set TX FIFO threshold and enable DMA. */
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
v |= ((sc->sc_variant == GEM_SUN_ERI ? 0x100 : 0x4ff) << 10) &
GEM_TX_CONFIG_TXFIFO_TH;
GEM_BANK1_WRITE_4(sc, GEM_TX_CONFIG, v | GEM_TX_CONFIG_TXDMA_EN);
/* step 10. ERX Configuration */
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
/* Encode Receive Descriptor ring size. */
v = gem_ringsize(GEM_NRXDESC /* XXX */);
/* RX TCP/UDP checksum offset */
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
v |= ((ETHER_HDR_LEN + sizeof(struct ip)) <<
GEM_RX_CONFIG_CXM_START_SHFT);
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
/* Set RX FIFO threshold, set first byte offset and enable DMA. */
GEM_BANK1_WRITE_4(sc, GEM_RX_CONFIG,
v | (GEM_THRSH_1024 << GEM_RX_CONFIG_FIFO_THRS_SHIFT) |
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
(ETHER_ALIGN << GEM_RX_CONFIG_FBOFF_SHFT) |
GEM_RX_CONFIG_RXDMA_EN);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
/* Adjusting for the SBus clock probably isn't worth the fuzz. */
GEM_BANK1_WRITE_4(sc, GEM_RX_BLANKING,
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
((6 * (sc->sc_flags & GEM_PCI66) != 0 ? 2 : 1) <<
GEM_RX_BLANKING_TIME_SHIFT) | 6);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
/*
* The following value is for an OFF Threshold of about 3/4 full
* and an ON Threshold of 1/4 full.
*/
GEM_BANK1_WRITE_4(sc, GEM_RX_PAUSE_THRESH,
(3 * sc->sc_rxfifosize / 256) |
((sc->sc_rxfifosize / 256) << 12));
/* step 11. Configure Media. */
/* step 12. RX_MAC Configuration Register */
v = GEM_BANK1_READ_4(sc, GEM_MAC_RX_CONFIG);
v &= ~GEM_MAC_RX_ENABLE;
v |= GEM_MAC_RX_STRIP_CRC;
sc->sc_mac_rxcfg = v;
/*
* Clear the RX filter and reprogram it. This will also set the
* current RX MAC configuration and enable it.
*/
gem_setladrf(sc);
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
/* step 13. TX_MAC Configuration Register */
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
v = GEM_BANK1_READ_4(sc, GEM_MAC_TX_CONFIG);
v |= GEM_MAC_TX_ENABLE;
(void)gem_disable_tx(sc);
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
GEM_BANK1_WRITE_4(sc, GEM_MAC_TX_CONFIG, v);
/* step 14. Issue Transmit Pending command. */
/* step 15. Give the receiver a swift kick. */
GEM_BANK1_WRITE_4(sc, GEM_RX_KICK, GEM_NRXDESC - 4);
ifp->if_drv_flags |= IFF_DRV_RUNNING;
ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
mii_mediachg(sc->sc_mii);
/* Start the one second timer. */
sc->sc_wdog_timer = 0;
callout_reset(&sc->sc_tick_ch, hz, gem_tick, sc);
}
static int
gem_load_txmbuf(struct gem_softc *sc, struct mbuf **m_head)
{
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
bus_dma_segment_t txsegs[GEM_NTXSEGS];
struct gem_txsoft *txs;
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
struct ip *ip;
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
struct mbuf *m;
uint64_t cflags, flags;
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
int error, nexttx, nsegs, offset, seg;
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
GEM_LOCK_ASSERT(sc, MA_OWNED);
/* Get a work queue entry. */
if ((txs = STAILQ_FIRST(&sc->sc_txfreeq)) == NULL) {
/* Ran out of descriptors. */
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
return (ENOBUFS);
}
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
cflags = 0;
if (((*m_head)->m_pkthdr.csum_flags & sc->sc_csum_features) != 0) {
if (M_WRITABLE(*m_head) == 0) {
m = m_dup(*m_head, M_NOWAIT);
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
m_freem(*m_head);
*m_head = m;
if (m == NULL)
return (ENOBUFS);
}
offset = sizeof(struct ether_header);
m = m_pullup(*m_head, offset + sizeof(struct ip));
if (m == NULL) {
*m_head = NULL;
return (ENOBUFS);
}
ip = (struct ip *)(mtod(m, caddr_t) + offset);
offset += (ip->ip_hl << 2);
cflags = offset << GEM_TD_CXSUM_STARTSHFT |
((offset + m->m_pkthdr.csum_data) <<
GEM_TD_CXSUM_STUFFSHFT) | GEM_TD_CXSUM_ENABLE;
*m_head = m;
}
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
error = bus_dmamap_load_mbuf_sg(sc->sc_tdmatag, txs->txs_dmamap,
*m_head, txsegs, &nsegs, BUS_DMA_NOWAIT);
if (error == EFBIG) {
m = m_collapse(*m_head, M_NOWAIT, GEM_NTXSEGS);
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
if (m == NULL) {
m_freem(*m_head);
*m_head = NULL;
return (ENOBUFS);
}
*m_head = m;
error = bus_dmamap_load_mbuf_sg(sc->sc_tdmatag,
txs->txs_dmamap, *m_head, txsegs, &nsegs,
BUS_DMA_NOWAIT);
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
if (error != 0) {
m_freem(*m_head);
*m_head = NULL;
return (error);
}
} else if (error != 0)
return (error);
/* If nsegs is wrong then the stack is corrupt. */
KASSERT(nsegs <= GEM_NTXSEGS,
("%s: too many DMA segments (%d)", __func__, nsegs));
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
if (nsegs == 0) {
m_freem(*m_head);
*m_head = NULL;
return (EIO);
}
/*
* Ensure we have enough descriptors free to describe
* the packet. Note, we always reserve one descriptor
* at the end of the ring as a termination point, in
* order to prevent wrap-around.
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
*/
if (nsegs > sc->sc_txfree - 1) {
txs->txs_ndescs = 0;
bus_dmamap_unload(sc->sc_tdmatag, txs->txs_dmamap);
return (ENOBUFS);
}
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
txs->txs_ndescs = nsegs;
txs->txs_firstdesc = sc->sc_txnext;
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
nexttx = txs->txs_firstdesc;
for (seg = 0; seg < nsegs; seg++, nexttx = GEM_NEXTTX(nexttx)) {
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
#ifdef GEM_DEBUG
CTR6(KTR_GEM,
"%s: mapping seg %d (txd %d), len %lx, addr %#lx (%#lx)",
__func__, seg, nexttx, txsegs[seg].ds_len,
txsegs[seg].ds_addr,
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
GEM_DMA_WRITE(sc, txsegs[seg].ds_addr));
#endif
sc->sc_txdescs[nexttx].gd_addr =
GEM_DMA_WRITE(sc, txsegs[seg].ds_addr);
KASSERT(txsegs[seg].ds_len < GEM_TD_BUFSIZE,
("%s: segment size too large!", __func__));
flags = txsegs[seg].ds_len & GEM_TD_BUFSIZE;
sc->sc_txdescs[nexttx].gd_flags =
GEM_DMA_WRITE(sc, flags | cflags);
txs->txs_lastdesc = nexttx;
}
/* Set EOP on the last descriptor. */
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
#ifdef GEM_DEBUG
CTR3(KTR_GEM, "%s: end of packet at segment %d, TX %d",
__func__, seg, nexttx);
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
#endif
sc->sc_txdescs[txs->txs_lastdesc].gd_flags |=
GEM_DMA_WRITE(sc, GEM_TD_END_OF_PACKET);
/* Lastly set SOP on the first descriptor. */
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
#ifdef GEM_DEBUG
CTR3(KTR_GEM, "%s: start of packet at segment %d, TX %d",
__func__, seg, nexttx);
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
#endif
if (++sc->sc_txwin > GEM_NTXSEGS * 2 / 3) {
sc->sc_txwin = 0;
sc->sc_txdescs[txs->txs_firstdesc].gd_flags |=
GEM_DMA_WRITE(sc, GEM_TD_INTERRUPT_ME |
GEM_TD_START_OF_PACKET);
} else
sc->sc_txdescs[txs->txs_firstdesc].gd_flags |=
GEM_DMA_WRITE(sc, GEM_TD_START_OF_PACKET);
/* Sync the DMA map. */
bus_dmamap_sync(sc->sc_tdmatag, txs->txs_dmamap,
BUS_DMASYNC_PREWRITE);
#ifdef GEM_DEBUG
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
CTR4(KTR_GEM, "%s: setting firstdesc=%d, lastdesc=%d, ndescs=%d",
__func__, txs->txs_firstdesc, txs->txs_lastdesc,
txs->txs_ndescs);
#endif
STAILQ_REMOVE_HEAD(&sc->sc_txfreeq, txs_q);
STAILQ_INSERT_TAIL(&sc->sc_txdirtyq, txs, txs_q);
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
txs->txs_mbuf = *m_head;
sc->sc_txnext = GEM_NEXTTX(txs->txs_lastdesc);
sc->sc_txfree -= txs->txs_ndescs;
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
return (0);
}
static void
gem_init_regs(struct gem_softc *sc)
{
const u_char *laddr = IF_LLADDR(sc->sc_ifp);
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
GEM_LOCK_ASSERT(sc, MA_OWNED);
/* These registers are not cleared on reset. */
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
if ((sc->sc_flags & GEM_INITED) == 0) {
/* magic values */
GEM_BANK1_WRITE_4(sc, GEM_MAC_IPG0, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_IPG1, 8);
GEM_BANK1_WRITE_4(sc, GEM_MAC_IPG2, 4);
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
/* min frame length */
GEM_BANK1_WRITE_4(sc, GEM_MAC_MAC_MIN_FRAME, ETHER_MIN_LEN);
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
/* max frame length and max burst size */
GEM_BANK1_WRITE_4(sc, GEM_MAC_MAC_MAX_FRAME,
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
(ETHER_MAX_LEN + ETHER_VLAN_ENCAP_LEN) | (0x2000 << 16));
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
/* more magic values */
GEM_BANK1_WRITE_4(sc, GEM_MAC_PREAMBLE_LEN, 0x7);
GEM_BANK1_WRITE_4(sc, GEM_MAC_JAM_SIZE, 0x4);
GEM_BANK1_WRITE_4(sc, GEM_MAC_ATTEMPT_LIMIT, 0x10);
GEM_BANK1_WRITE_4(sc, GEM_MAC_CONTROL_TYPE, 0x8808);
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
/* random number seed */
GEM_BANK1_WRITE_4(sc, GEM_MAC_RANDOM_SEED,
((laddr[5] << 8) | laddr[4]) & 0x3ff);
/* secondary MAC address: 0:0:0:0:0:0 */
GEM_BANK1_WRITE_4(sc, GEM_MAC_ADDR3, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_ADDR4, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_ADDR5, 0);
/* MAC control address: 01:80:c2:00:00:01 */
GEM_BANK1_WRITE_4(sc, GEM_MAC_ADDR6, 0x0001);
GEM_BANK1_WRITE_4(sc, GEM_MAC_ADDR7, 0xc200);
GEM_BANK1_WRITE_4(sc, GEM_MAC_ADDR8, 0x0180);
/* MAC filter address: 0:0:0:0:0:0 */
GEM_BANK1_WRITE_4(sc, GEM_MAC_ADDR_FILTER0, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_ADDR_FILTER1, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_ADDR_FILTER2, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_ADR_FLT_MASK1_2, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_ADR_FLT_MASK0, 0);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
sc->sc_flags |= GEM_INITED;
}
/* Counters need to be zeroed. */
GEM_BANK1_WRITE_4(sc, GEM_MAC_NORM_COLL_CNT, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_FIRST_COLL_CNT, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_EXCESS_COLL_CNT, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_LATE_COLL_CNT, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_DEFER_TMR_CNT, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_PEAK_ATTEMPTS, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_RX_FRAME_COUNT, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_RX_LEN_ERR_CNT, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_RX_ALIGN_ERR, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_RX_CRC_ERR_CNT, 0);
GEM_BANK1_WRITE_4(sc, GEM_MAC_RX_CODE_VIOL, 0);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
/* Set XOFF PAUSE time. */
GEM_BANK1_WRITE_4(sc, GEM_MAC_SEND_PAUSE_CMD, 0x1BF0);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
/* Set the station address. */
GEM_BANK1_WRITE_4(sc, GEM_MAC_ADDR0, (laddr[4] << 8) | laddr[5]);
GEM_BANK1_WRITE_4(sc, GEM_MAC_ADDR1, (laddr[2] << 8) | laddr[3]);
GEM_BANK1_WRITE_4(sc, GEM_MAC_ADDR2, (laddr[0] << 8) | laddr[1]);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
/* Enable MII outputs. */
GEM_BANK1_WRITE_4(sc, GEM_MAC_XIF_CONFIG, GEM_MAC_XIF_TX_MII_ENA);
}
static void
gem_start(struct ifnet *ifp)
{
struct gem_softc *sc = ifp->if_softc;
GEM_LOCK(sc);
gem_start_locked(ifp);
GEM_UNLOCK(sc);
}
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
static inline void
gem_txkick(struct gem_softc *sc)
{
/*
* Update the TX kick register. This register has to point to the
* descriptor after the last valid one and for optimum performance
* should be incremented in multiples of 4 (the DMA engine fetches/
* updates descriptors in batches of 4).
*/
#ifdef GEM_DEBUG
CTR3(KTR_GEM, "%s: %s: kicking TX %d",
device_get_name(sc->sc_dev), __func__, sc->sc_txnext);
#endif
GEM_CDSYNC(sc, BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
GEM_BANK1_WRITE_4(sc, GEM_TX_KICK, sc->sc_txnext);
}
static void
gem_start_locked(struct ifnet *ifp)
{
struct gem_softc *sc = ifp->if_softc;
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
struct mbuf *m;
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
int kicked, ntx;
GEM_LOCK_ASSERT(sc, MA_OWNED);
if ((ifp->if_drv_flags & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
IFF_DRV_RUNNING || (sc->sc_flags & GEM_LINK) == 0)
return;
#ifdef GEM_DEBUG
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
CTR4(KTR_GEM, "%s: %s: txfree %d, txnext %d",
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
device_get_name(sc->sc_dev), __func__, sc->sc_txfree,
sc->sc_txnext);
#endif
ntx = 0;
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
kicked = 0;
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
for (; !IFQ_DRV_IS_EMPTY(&ifp->if_snd) && sc->sc_txfree > 1;) {
IFQ_DRV_DEQUEUE(&ifp->if_snd, m);
if (m == NULL)
break;
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
if (gem_load_txmbuf(sc, &m) != 0) {
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
if (m == NULL)
break;
ifp->if_drv_flags |= IFF_DRV_OACTIVE;
IFQ_DRV_PREPEND(&ifp->if_snd, m);
break;
}
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
if ((sc->sc_txnext % 4) == 0) {
gem_txkick(sc);
kicked = 1;
} else
kicked = 0;
ntx++;
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
BPF_MTAP(ifp, m);
}
if (ntx > 0) {
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
if (kicked == 0)
gem_txkick(sc);
#ifdef GEM_DEBUG
CTR2(KTR_GEM, "%s: packets enqueued, OWN on %d",
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
device_get_name(sc->sc_dev), sc->sc_txnext);
#endif
/* Set a watchdog timer in case the chip flakes out. */
sc->sc_wdog_timer = 5;
#ifdef GEM_DEBUG
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
CTR3(KTR_GEM, "%s: %s: watchdog %d",
device_get_name(sc->sc_dev), __func__,
sc->sc_wdog_timer);
#endif
}
}
static void
gem_tint(struct gem_softc *sc)
{
struct ifnet *ifp = sc->sc_ifp;
struct gem_txsoft *txs;
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
int progress;
uint32_t txlast;
#ifdef GEM_DEBUG
int i;
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
GEM_LOCK_ASSERT(sc, MA_OWNED);
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
CTR2(KTR_GEM, "%s: %s", device_get_name(sc->sc_dev), __func__);
#endif
/*
* Go through our TX list and free mbufs for those
* frames that have been transmitted.
*/
progress = 0;
GEM_CDSYNC(sc, BUS_DMASYNC_POSTREAD);
while ((txs = STAILQ_FIRST(&sc->sc_txdirtyq)) != NULL) {
#ifdef GEM_DEBUG
if ((ifp->if_flags & IFF_DEBUG) != 0) {
printf(" txsoft %p transmit chain:\n", txs);
for (i = txs->txs_firstdesc;; i = GEM_NEXTTX(i)) {
printf("descriptor %d: ", i);
printf("gd_flags: 0x%016llx\t",
(long long)GEM_DMA_READ(sc,
sc->sc_txdescs[i].gd_flags));
printf("gd_addr: 0x%016llx\n",
(long long)GEM_DMA_READ(sc,
sc->sc_txdescs[i].gd_addr));
if (i == txs->txs_lastdesc)
break;
}
}
#endif
/*
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
* In theory, we could harvest some descriptors before
* the ring is empty, but that's a bit complicated.
*
* GEM_TX_COMPLETION points to the last descriptor
* processed + 1.
*/
txlast = GEM_BANK1_READ_4(sc, GEM_TX_COMPLETION);
#ifdef GEM_DEBUG
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
CTR4(KTR_GEM, "%s: txs->txs_firstdesc = %d, "
"txs->txs_lastdesc = %d, txlast = %d",
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
__func__, txs->txs_firstdesc, txs->txs_lastdesc, txlast);
#endif
if (txs->txs_firstdesc <= txs->txs_lastdesc) {
if ((txlast >= txs->txs_firstdesc) &&
(txlast <= txs->txs_lastdesc))
break;
} else {
/* Ick -- this command wraps. */
if ((txlast >= txs->txs_firstdesc) ||
(txlast <= txs->txs_lastdesc))
break;
}
#ifdef GEM_DEBUG
CTR1(KTR_GEM, "%s: releasing a descriptor", __func__);
#endif
STAILQ_REMOVE_HEAD(&sc->sc_txdirtyq, txs_q);
sc->sc_txfree += txs->txs_ndescs;
bus_dmamap_sync(sc->sc_tdmatag, txs->txs_dmamap,
BUS_DMASYNC_POSTWRITE);
bus_dmamap_unload(sc->sc_tdmatag, txs->txs_dmamap);
if (txs->txs_mbuf != NULL) {
m_freem(txs->txs_mbuf);
txs->txs_mbuf = NULL;
}
STAILQ_INSERT_TAIL(&sc->sc_txfreeq, txs, txs_q);
if_inc_counter(ifp, IFCOUNTER_OPACKETS, 1);
progress = 1;
}
#ifdef GEM_DEBUG
CTR4(KTR_GEM, "%s: GEM_TX_STATE_MACHINE %x GEM_TX_DATA_PTR %llx "
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
"GEM_TX_COMPLETION %x",
__func__, GEM_BANK1_READ_4(sc, GEM_TX_STATE_MACHINE),
((long long)GEM_BANK1_READ_4(sc, GEM_TX_DATA_PTR_HI) << 32) |
GEM_BANK1_READ_4(sc, GEM_TX_DATA_PTR_LO),
GEM_BANK1_READ_4(sc, GEM_TX_COMPLETION));
#endif
if (progress) {
if (sc->sc_txfree == GEM_NTXDESC - 1)
sc->sc_txwin = 0;
/*
* We freed some descriptors, so reset IFF_DRV_OACTIVE
* and restart.
*/
ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
if (STAILQ_EMPTY(&sc->sc_txdirtyq))
sc->sc_wdog_timer = 0;
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
gem_start_locked(ifp);
}
#ifdef GEM_DEBUG
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
CTR3(KTR_GEM, "%s: %s: watchdog %d",
device_get_name(sc->sc_dev), __func__, sc->sc_wdog_timer);
#endif
}
#ifdef GEM_RINT_TIMEOUT
static void
gem_rint_timeout(void *arg)
{
struct gem_softc *sc = arg;
GEM_LOCK_ASSERT(sc, MA_OWNED);
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
gem_rint(sc);
}
2002-07-24 01:09:38 +00:00
#endif
static void
gem_rint(struct gem_softc *sc)
{
struct ifnet *ifp = sc->sc_ifp;
struct mbuf *m;
uint64_t rxstat;
uint32_t rxcomp;
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
GEM_LOCK_ASSERT(sc, MA_OWNED);
#ifdef GEM_RINT_TIMEOUT
callout_stop(&sc->sc_rx_ch);
#endif
#ifdef GEM_DEBUG
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
CTR2(KTR_GEM, "%s: %s", device_get_name(sc->sc_dev), __func__);
#endif
/*
* Read the completion register once. This limits
* how long the following loop can execute.
*/
rxcomp = GEM_BANK1_READ_4(sc, GEM_RX_COMPLETION);
#ifdef GEM_DEBUG
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
CTR3(KTR_GEM, "%s: sc->sc_rxptr %d, complete %d",
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
__func__, sc->sc_rxptr, rxcomp);
#endif
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
GEM_CDSYNC(sc, BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
for (; sc->sc_rxptr != rxcomp;) {
m = sc->sc_rxsoft[sc->sc_rxptr].rxs_mbuf;
rxstat = GEM_DMA_READ(sc,
sc->sc_rxdescs[sc->sc_rxptr].gd_flags);
if (rxstat & GEM_RD_OWN) {
#ifdef GEM_RINT_TIMEOUT
/*
* The descriptor is still marked as owned, although
* it is supposed to have completed. This has been
* observed on some machines. Just exiting here
* might leave the packet sitting around until another
* one arrives to trigger a new interrupt, which is
* generally undesirable, so set up a timeout.
*/
callout_reset(&sc->sc_rx_ch, GEM_RXOWN_TICKS,
gem_rint_timeout, sc);
#endif
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
m = NULL;
goto kickit;
}
if (rxstat & GEM_RD_BAD_CRC) {
if_inc_counter(ifp, IFCOUNTER_IERRORS, 1);
device_printf(sc->sc_dev, "receive error: CRC error\n");
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
GEM_INIT_RXDESC(sc, sc->sc_rxptr);
m = NULL;
goto kickit;
}
#ifdef GEM_DEBUG
if ((ifp->if_flags & IFF_DEBUG) != 0) {
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
printf(" rxsoft %p descriptor %d: ",
&sc->sc_rxsoft[sc->sc_rxptr], sc->sc_rxptr);
printf("gd_flags: 0x%016llx\t",
(long long)GEM_DMA_READ(sc,
sc->sc_rxdescs[sc->sc_rxptr].gd_flags));
printf("gd_addr: 0x%016llx\n",
(long long)GEM_DMA_READ(sc,
sc->sc_rxdescs[sc->sc_rxptr].gd_addr));
}
#endif
/*
* Allocate a new mbuf cluster. If that fails, we are
* out of memory, and must drop the packet and recycle
* the buffer that's already attached to this descriptor.
*/
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
if (gem_add_rxbuf(sc, sc->sc_rxptr) != 0) {
if_inc_counter(ifp, IFCOUNTER_IQDROPS, 1);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
GEM_INIT_RXDESC(sc, sc->sc_rxptr);
m = NULL;
}
kickit:
/*
* Update the RX kick register. This register has to point
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
* to the descriptor after the last valid one (before the
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
* current batch) and for optimum performance should be
* incremented in multiples of 4 (the DMA engine fetches/
* updates descriptors in batches of 4).
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
*/
sc->sc_rxptr = GEM_NEXTRX(sc->sc_rxptr);
if ((sc->sc_rxptr % 4) == 0) {
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
GEM_CDSYNC(sc,
BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
GEM_BANK1_WRITE_4(sc, GEM_RX_KICK,
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
(sc->sc_rxptr + GEM_NRXDESC - 4) &
GEM_NRXDESC_MASK);
}
if (m == NULL) {
if (rxstat & GEM_RD_OWN)
break;
continue;
}
if_inc_counter(ifp, IFCOUNTER_IPACKETS, 1);
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
m->m_data += ETHER_ALIGN; /* first byte offset */
m->m_pkthdr.rcvif = ifp;
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
m->m_pkthdr.len = m->m_len = GEM_RD_BUFLEN(rxstat);
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
if ((ifp->if_capenable & IFCAP_RXCSUM) != 0)
gem_rxcksum(m, rxstat);
/* Pass it on. */
GEM_UNLOCK(sc);
(*ifp->if_input)(ifp, m);
GEM_LOCK(sc);
}
#ifdef GEM_DEBUG
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
CTR3(KTR_GEM, "%s: done sc->sc_rxptr %d, complete %d", __func__,
sc->sc_rxptr, GEM_BANK1_READ_4(sc, GEM_RX_COMPLETION));
#endif
}
static int
gem_add_rxbuf(struct gem_softc *sc, int idx)
{
struct gem_rxsoft *rxs = &sc->sc_rxsoft[idx];
struct mbuf *m;
bus_dma_segment_t segs[1];
int error, nsegs;
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
GEM_LOCK_ASSERT(sc, MA_OWNED);
m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
if (m == NULL)
return (ENOBUFS);
m->m_len = m->m_pkthdr.len = m->m_ext.ext_size;
#ifdef GEM_DEBUG
/* Bzero the packet to check DMA. */
memset(m->m_ext.ext_buf, 0, m->m_ext.ext_size);
#endif
if (rxs->rxs_mbuf != NULL) {
bus_dmamap_sync(sc->sc_rdmatag, rxs->rxs_dmamap,
BUS_DMASYNC_POSTREAD);
bus_dmamap_unload(sc->sc_rdmatag, rxs->rxs_dmamap);
}
error = bus_dmamap_load_mbuf_sg(sc->sc_rdmatag, rxs->rxs_dmamap,
m, segs, &nsegs, BUS_DMA_NOWAIT);
if (error != 0) {
device_printf(sc->sc_dev,
"cannot load RS DMA map %d, error = %d\n", idx, error);
m_freem(m);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
return (error);
}
/* If nsegs is wrong then the stack is corrupt. */
KASSERT(nsegs == 1,
("%s: too many DMA segments (%d)", __func__, nsegs));
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
rxs->rxs_mbuf = m;
rxs->rxs_paddr = segs[0].ds_addr;
bus_dmamap_sync(sc->sc_rdmatag, rxs->rxs_dmamap,
BUS_DMASYNC_PREREAD);
GEM_INIT_RXDESC(sc, idx);
return (0);
}
static void
gem_eint(struct gem_softc *sc, u_int status)
{
if_inc_counter(sc->sc_ifp, IFCOUNTER_IERRORS, 1);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
if ((status & GEM_INTR_RX_TAG_ERR) != 0) {
gem_reset_rxdma(sc);
return;
}
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
device_printf(sc->sc_dev, "%s: status 0x%x", __func__, status);
if ((status & GEM_INTR_BERR) != 0) {
if ((sc->sc_flags & GEM_PCI) != 0)
printf(", PCI bus error 0x%x\n",
GEM_BANK1_READ_4(sc, GEM_PCI_ERROR_STATUS));
else
printf(", SBus error 0x%x\n",
GEM_BANK1_READ_4(sc, GEM_SBUS_STATUS));
}
}
void
gem_intr(void *v)
{
struct gem_softc *sc = v;
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
uint32_t status, status2;
GEM_LOCK(sc);
status = GEM_BANK1_READ_4(sc, GEM_STATUS);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
#ifdef GEM_DEBUG
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
CTR4(KTR_GEM, "%s: %s: cplt %x, status %x",
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
device_get_name(sc->sc_dev), __func__,
(status >> GEM_STATUS_TX_COMPLETION_SHFT), (u_int)status);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
/*
* PCS interrupts must be cleared, otherwise no traffic is passed!
*/
if ((status & GEM_INTR_PCS) != 0) {
status2 =
GEM_BANK1_READ_4(sc, GEM_MII_INTERRUP_STATUS) |
GEM_BANK1_READ_4(sc, GEM_MII_INTERRUP_STATUS);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
if ((status2 & GEM_MII_INTERRUP_LINK) != 0)
device_printf(sc->sc_dev,
"%s: PCS link status changed\n", __func__);
}
if ((status & GEM_MAC_CONTROL_STATUS) != 0) {
status2 = GEM_BANK1_READ_4(sc, GEM_MAC_CONTROL_STATUS);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
if ((status2 & GEM_MAC_PAUSED) != 0)
device_printf(sc->sc_dev,
"%s: PAUSE received (PAUSE time %d slots)\n",
__func__, GEM_MAC_PAUSE_TIME(status2));
if ((status2 & GEM_MAC_PAUSE) != 0)
device_printf(sc->sc_dev,
"%s: transited to PAUSE state\n", __func__);
if ((status2 & GEM_MAC_RESUME) != 0)
device_printf(sc->sc_dev,
"%s: transited to non-PAUSE state\n", __func__);
}
if ((status & GEM_INTR_MIF) != 0)
device_printf(sc->sc_dev, "%s: MIF interrupt\n", __func__);
#endif
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
if (__predict_false(status &
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
(GEM_INTR_RX_TAG_ERR | GEM_INTR_PERR | GEM_INTR_BERR)) != 0)
gem_eint(sc, status);
if ((status & (GEM_INTR_RX_DONE | GEM_INTR_RX_NOBUF)) != 0)
gem_rint(sc);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
if ((status & (GEM_INTR_TX_EMPTY | GEM_INTR_TX_INTME)) != 0)
gem_tint(sc);
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
if (__predict_false((status & GEM_INTR_TX_MAC) != 0)) {
status2 = GEM_BANK1_READ_4(sc, GEM_MAC_TX_STATUS);
if ((status2 &
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
~(GEM_MAC_TX_XMIT_DONE | GEM_MAC_TX_DEFER_EXP |
GEM_MAC_TX_PEAK_EXP)) != 0)
device_printf(sc->sc_dev,
"MAC TX fault, status %x\n", status2);
if ((status2 &
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
(GEM_MAC_TX_UNDERRUN | GEM_MAC_TX_PKT_TOO_LONG)) != 0) {
if_inc_counter(sc->sc_ifp, IFCOUNTER_OERRORS, 1);
sc->sc_ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
gem_init_locked(sc);
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
}
}
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
if (__predict_false((status & GEM_INTR_RX_MAC) != 0)) {
status2 = GEM_BANK1_READ_4(sc, GEM_MAC_RX_STATUS);
/*
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
* At least with GEM_SUN_GEM and some GEM_SUN_ERI
* revisions GEM_MAC_RX_OVERFLOW happen often due to a
* silicon bug so handle them silently. Moreover, it's
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
* likely that the receiver has hung so we reset it.
*/
if ((status2 & GEM_MAC_RX_OVERFLOW) != 0) {
if_inc_counter(sc->sc_ifp, IFCOUNTER_IERRORS, 1);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
gem_reset_rxdma(sc);
} else if ((status2 &
~(GEM_MAC_RX_DONE | GEM_MAC_RX_FRAME_CNT)) != 0)
device_printf(sc->sc_dev,
"MAC RX fault, status %x\n", status2);
}
GEM_UNLOCK(sc);
}
static int
gem_watchdog(struct gem_softc *sc)
{
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
struct ifnet *ifp = sc->sc_ifp;
GEM_LOCK_ASSERT(sc, MA_OWNED);
#ifdef GEM_DEBUG
CTR4(KTR_GEM,
"%s: GEM_RX_CONFIG %x GEM_MAC_RX_STATUS %x GEM_MAC_RX_CONFIG %x",
__func__, GEM_BANK1_READ_4(sc, GEM_RX_CONFIG),
GEM_BANK1_READ_4(sc, GEM_MAC_RX_STATUS),
GEM_BANK1_READ_4(sc, GEM_MAC_RX_CONFIG));
CTR4(KTR_GEM,
"%s: GEM_TX_CONFIG %x GEM_MAC_TX_STATUS %x GEM_MAC_TX_CONFIG %x",
__func__, GEM_BANK1_READ_4(sc, GEM_TX_CONFIG),
GEM_BANK1_READ_4(sc, GEM_MAC_TX_STATUS),
GEM_BANK1_READ_4(sc, GEM_MAC_TX_CONFIG));
#endif
if (sc->sc_wdog_timer == 0 || --sc->sc_wdog_timer != 0)
return (0);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
if ((sc->sc_flags & GEM_LINK) != 0)
device_printf(sc->sc_dev, "device timeout\n");
else if (bootverbose)
device_printf(sc->sc_dev, "device timeout (no link)\n");
if_inc_counter(ifp, IFCOUNTER_OERRORS, 1);
/* Try to get more packets going. */
ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
gem_init_locked(sc);
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
gem_start_locked(ifp);
return (EJUSTRETURN);
}
static void
gem_mifinit(struct gem_softc *sc)
{
/* Configure the MIF in frame mode. */
GEM_BANK1_WRITE_4(sc, GEM_MIF_CONFIG,
GEM_BANK1_READ_4(sc, GEM_MIF_CONFIG) & ~GEM_MIF_CONFIG_BB_ENA);
GEM_BANK1_BARRIER(sc, GEM_MIF_CONFIG, 4,
BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE);
}
/*
* MII interface
*
* The MII interface supports at least three different operating modes:
*
* Bitbang mode is implemented using data, clock and output enable registers.
*
* Frame mode is implemented by loading a complete frame into the frame
* register and polling the valid bit for completion.
*
* Polling mode uses the frame register but completion is indicated by
* an interrupt.
*
*/
int
gem_mii_readreg(device_t dev, int phy, int reg)
{
struct gem_softc *sc;
int n;
uint32_t v;
#ifdef GEM_DEBUG_PHY
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
printf("%s: phy %d reg %d\n", __func__, phy, reg);
#endif
sc = device_get_softc(dev);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
if ((sc->sc_flags & GEM_SERDES) != 0) {
switch (reg) {
case MII_BMCR:
reg = GEM_MII_CONTROL;
break;
case MII_BMSR:
reg = GEM_MII_STATUS;
break;
case MII_PHYIDR1:
case MII_PHYIDR2:
return (0);
case MII_ANAR:
reg = GEM_MII_ANAR;
break;
case MII_ANLPAR:
reg = GEM_MII_ANLPAR;
break;
case MII_EXTSR:
return (EXTSR_1000XFDX | EXTSR_1000XHDX);
default:
device_printf(sc->sc_dev,
"%s: unhandled register %d\n", __func__, reg);
return (0);
}
return (GEM_BANK1_READ_4(sc, reg));
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
}
/* Construct the frame command. */
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
v = GEM_MIF_FRAME_READ |
(phy << GEM_MIF_PHY_SHIFT) |
(reg << GEM_MIF_REG_SHIFT);
GEM_BANK1_WRITE_4(sc, GEM_MIF_FRAME, v);
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
GEM_BANK1_BARRIER(sc, GEM_MIF_FRAME, 4,
BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE);
for (n = 0; n < 100; n++) {
DELAY(1);
v = GEM_BANK1_READ_4(sc, GEM_MIF_FRAME);
if (v & GEM_MIF_FRAME_TA0)
return (v & GEM_MIF_FRAME_DATA);
}
device_printf(sc->sc_dev, "%s: timed out\n", __func__);
return (0);
}
int
gem_mii_writereg(device_t dev, int phy, int reg, int val)
{
struct gem_softc *sc;
int n;
uint32_t v;
#ifdef GEM_DEBUG_PHY
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
printf("%s: phy %d reg %d val %x\n", phy, reg, val, __func__);
#endif
sc = device_get_softc(dev);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
if ((sc->sc_flags & GEM_SERDES) != 0) {
switch (reg) {
case MII_BMSR:
reg = GEM_MII_STATUS;
break;
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
case MII_BMCR:
reg = GEM_MII_CONTROL;
if ((val & GEM_MII_CONTROL_RESET) == 0)
break;
GEM_BANK1_WRITE_4(sc, GEM_MII_CONTROL, val);
GEM_BANK1_BARRIER(sc, GEM_MII_CONTROL, 4,
BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE);
if (!GEM_BANK1_BITWAIT(sc, GEM_MII_CONTROL,
GEM_MII_CONTROL_RESET, 0))
device_printf(sc->sc_dev,
"cannot reset PCS\n");
/* FALLTHROUGH */
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
case MII_ANAR:
GEM_BANK1_WRITE_4(sc, GEM_MII_CONFIG, 0);
GEM_BANK1_BARRIER(sc, GEM_MII_CONFIG, 4,
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
BUS_SPACE_BARRIER_WRITE);
GEM_BANK1_WRITE_4(sc, GEM_MII_ANAR, val);
GEM_BANK1_BARRIER(sc, GEM_MII_ANAR, 4,
BUS_SPACE_BARRIER_WRITE);
GEM_BANK1_WRITE_4(sc, GEM_MII_SLINK_CONTROL,
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
GEM_MII_SLINK_LOOPBACK | GEM_MII_SLINK_EN_SYNC_D);
GEM_BANK1_BARRIER(sc, GEM_MII_SLINK_CONTROL, 4,
BUS_SPACE_BARRIER_WRITE);
GEM_BANK1_WRITE_4(sc, GEM_MII_CONFIG,
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
GEM_MII_CONFIG_ENABLE);
GEM_BANK1_BARRIER(sc, GEM_MII_CONFIG, 4,
BUS_SPACE_BARRIER_WRITE);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
return (0);
case MII_ANLPAR:
reg = GEM_MII_ANLPAR;
break;
default:
device_printf(sc->sc_dev,
"%s: unhandled register %d\n", __func__, reg);
return (0);
}
GEM_BANK1_WRITE_4(sc, reg, val);
GEM_BANK1_BARRIER(sc, reg, 4,
BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
return (0);
}
/* Construct the frame command. */
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
v = GEM_MIF_FRAME_WRITE |
(phy << GEM_MIF_PHY_SHIFT) |
(reg << GEM_MIF_REG_SHIFT) |
(val & GEM_MIF_FRAME_DATA);
GEM_BANK1_WRITE_4(sc, GEM_MIF_FRAME, v);
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
GEM_BANK1_BARRIER(sc, GEM_MIF_FRAME, 4,
BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE);
for (n = 0; n < 100; n++) {
DELAY(1);
v = GEM_BANK1_READ_4(sc, GEM_MIF_FRAME);
if (v & GEM_MIF_FRAME_TA0)
return (1);
}
device_printf(sc->sc_dev, "%s: timed out\n", __func__);
return (0);
}
void
gem_mii_statchg(device_t dev)
{
struct gem_softc *sc;
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
int gigabit;
uint32_t rxcfg, txcfg, v;
sc = device_get_softc(dev);
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
GEM_LOCK_ASSERT(sc, MA_OWNED);
#ifdef GEM_DEBUG
if ((sc->sc_ifp->if_flags & IFF_DEBUG) != 0)
device_printf(sc->sc_dev, "%s: status change\n", __func__);
#endif
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
if ((sc->sc_mii->mii_media_status & IFM_ACTIVE) != 0 &&
IFM_SUBTYPE(sc->sc_mii->mii_media_active) != IFM_NONE)
sc->sc_flags |= GEM_LINK;
else
sc->sc_flags &= ~GEM_LINK;
switch (IFM_SUBTYPE(sc->sc_mii->mii_media_active)) {
case IFM_1000_SX:
case IFM_1000_LX:
case IFM_1000_CX:
case IFM_1000_T:
gigabit = 1;
break;
default:
gigabit = 0;
}
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
/*
* The configuration done here corresponds to the steps F) and
* G) and as far as enabling of RX and TX MAC goes also step H)
* of the initialization sequence outlined in section 3.2.1 of
* the GEM Gigabit Ethernet ASIC Specification.
*/
rxcfg = sc->sc_mac_rxcfg;
rxcfg &= ~GEM_MAC_RX_CARR_EXTEND;
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
txcfg = GEM_MAC_TX_ENA_IPG0 | GEM_MAC_TX_NGU | GEM_MAC_TX_NGU_LIMIT;
if ((IFM_OPTIONS(sc->sc_mii->mii_media_active) & IFM_FDX) != 0)
txcfg |= GEM_MAC_TX_IGN_CARRIER | GEM_MAC_TX_IGN_COLLIS;
else if (gigabit != 0) {
rxcfg |= GEM_MAC_RX_CARR_EXTEND;
txcfg |= GEM_MAC_TX_CARR_EXTEND;
}
(void)gem_disable_tx(sc);
GEM_BANK1_WRITE_4(sc, GEM_MAC_TX_CONFIG, txcfg);
(void)gem_disable_rx(sc);
GEM_BANK1_WRITE_4(sc, GEM_MAC_RX_CONFIG, rxcfg);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
v = GEM_BANK1_READ_4(sc, GEM_MAC_CONTROL_CONFIG) &
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
~(GEM_MAC_CC_RX_PAUSE | GEM_MAC_CC_TX_PAUSE);
if ((IFM_OPTIONS(sc->sc_mii->mii_media_active) &
IFM_ETH_RXPAUSE) != 0)
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
v |= GEM_MAC_CC_RX_PAUSE;
if ((IFM_OPTIONS(sc->sc_mii->mii_media_active) &
IFM_ETH_TXPAUSE) != 0)
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
v |= GEM_MAC_CC_TX_PAUSE;
GEM_BANK1_WRITE_4(sc, GEM_MAC_CONTROL_CONFIG, v);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
if ((IFM_OPTIONS(sc->sc_mii->mii_media_active) & IFM_FDX) == 0 &&
gigabit != 0)
GEM_BANK1_WRITE_4(sc, GEM_MAC_SLOT_TIME,
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
GEM_MAC_SLOT_TIME_CARR_EXTEND);
else
GEM_BANK1_WRITE_4(sc, GEM_MAC_SLOT_TIME,
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
GEM_MAC_SLOT_TIME_NORMAL);
/* XIF Configuration */
v = GEM_MAC_XIF_LINK_LED;
v |= GEM_MAC_XIF_TX_MII_ENA;
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
if ((sc->sc_flags & GEM_SERDES) == 0) {
if ((GEM_BANK1_READ_4(sc, GEM_MIF_CONFIG) &
GEM_MIF_CONFIG_PHY_SEL) != 0) {
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
/* External MII needs echo disable if half duplex. */
- Initialize the ifnet structure, especially if_dname, before probing the PHYs as some PHY drivers use it (but probably shouldn't). How gem(4) has worked with brgphy(4) on powerpc without this so far is unclear to me. - Introduce a dying flag which is set during detach and checked in gem_ioctl() in order to prevent active BPF listeners to clear promiscuous mode which may lead to the tick callout being restarted which will trigger a panic once it's actually gone. - In gem_stop() reset rather than just disable the transmitter and receiver in order to ensure we're not unloading DMA maps still in use by the hardware. [1] - The blanking time is specified in PCI clocks so we should use twice the value when operating at 66MHz. - Spell some 2 as ETHER_ALIGN and a 19 as GEM_STATUS_TX_COMPLETION_SHFT to make the actual intentions clear. - As we don't unload the peak attempts counter ignore its overflow interrupts. - Remove a stale setting of a variable to GEM_TD_INTERRUPT_ME which isn't used afterwards. - For optimum performance increment the TX kick register in multiples of 4 if possible as suggested by the documentation. - Partially revert r164931; drivers should only clear the watchdog timer if all outstanding TX descriptors are done. - Fix some debugging strings. - Add a missing BUS_DMASYNC_POSTWRITE in gem_rint(). - As the error paths in the interrupt handler are generally unlikely predict them as false. - Add support for the SBus version of the GEM controller. [2] - Add some lock assertions. - Improve some comments. - Fix some more or less cosmetic issues in the code of the PCI front-end. - Change some softc members to be unsigned where more appropriate and remove unused ones. Approved by: re (kib) Obtained from: NetBSD (partially) [2], OpenBSD [1] MFC after: 2 weeks
2009-06-23 20:36:59 +00:00
if ((IFM_OPTIONS(sc->sc_mii->mii_media_active) &
IFM_FDX) == 0)
v |= GEM_MAC_XIF_ECHO_DISABL;
} else
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
/*
* Internal MII needs buffer enable.
* XXX buffer enable makes only sense for an
* external PHY.
*/
v |= GEM_MAC_XIF_MII_BUF_ENA;
}
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
if (gigabit != 0)
v |= GEM_MAC_XIF_GMII_MODE;
if ((IFM_OPTIONS(sc->sc_mii->mii_media_active) & IFM_FDX) != 0)
v |= GEM_MAC_XIF_FDPLX_LED;
GEM_BANK1_WRITE_4(sc, GEM_MAC_XIF_CONFIG, v);
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
sc->sc_mac_rxcfg = rxcfg;
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
if ((sc->sc_ifp->if_drv_flags & IFF_DRV_RUNNING) != 0 &&
(sc->sc_flags & GEM_LINK) != 0) {
GEM_BANK1_WRITE_4(sc, GEM_MAC_TX_CONFIG,
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
txcfg | GEM_MAC_TX_ENABLE);
GEM_BANK1_WRITE_4(sc, GEM_MAC_RX_CONFIG,
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
rxcfg | GEM_MAC_RX_ENABLE);
}
}
int
gem_mediachange(struct ifnet *ifp)
{
struct gem_softc *sc = ifp->if_softc;
int error;
/* XXX add support for serial media. */
GEM_LOCK(sc);
error = mii_mediachg(sc->sc_mii);
GEM_UNLOCK(sc);
return (error);
}
void
gem_mediastatus(struct ifnet *ifp, struct ifmediareq *ifmr)
{
struct gem_softc *sc = ifp->if_softc;
GEM_LOCK(sc);
if ((ifp->if_flags & IFF_UP) == 0) {
GEM_UNLOCK(sc);
return;
}
mii_pollstat(sc->sc_mii);
ifmr->ifm_active = sc->sc_mii->mii_media_active;
ifmr->ifm_status = sc->sc_mii->mii_media_status;
GEM_UNLOCK(sc);
}
static int
gem_ioctl(struct ifnet *ifp, u_long cmd, caddr_t data)
{
struct gem_softc *sc = ifp->if_softc;
struct ifreq *ifr = (struct ifreq *)data;
int error;
error = 0;
switch (cmd) {
case SIOCSIFFLAGS:
GEM_LOCK(sc);
if ((ifp->if_flags & IFF_UP) != 0) {
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
if ((ifp->if_drv_flags & IFF_DRV_RUNNING) != 0 &&
((ifp->if_flags ^ sc->sc_ifflags) &
(IFF_ALLMULTI | IFF_PROMISC)) != 0)
gem_setladrf(sc);
else
gem_init_locked(sc);
} else if ((ifp->if_drv_flags & IFF_DRV_RUNNING) != 0)
gem_stop(ifp, 0);
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
if ((ifp->if_flags & IFF_LINK0) != 0)
sc->sc_csum_features |= CSUM_UDP;
else
sc->sc_csum_features &= ~CSUM_UDP;
if ((ifp->if_capenable & IFCAP_TXCSUM) != 0)
ifp->if_hwassist = sc->sc_csum_features;
sc->sc_ifflags = ifp->if_flags;
GEM_UNLOCK(sc);
break;
case SIOCADDMULTI:
case SIOCDELMULTI:
GEM_LOCK(sc);
if ((ifp->if_drv_flags & IFF_DRV_RUNNING) != 0)
gem_setladrf(sc);
GEM_UNLOCK(sc);
break;
case SIOCGIFMEDIA:
case SIOCSIFMEDIA:
error = ifmedia_ioctl(ifp, ifr, &sc->sc_mii->mii_media, cmd);
break;
o Implemented Rx/Tx checksum offload. The simple checksum logic in GEMs is unable to discriminate UDP from TCP packets such that it can generate 0x0000 checksum value for the UDP datagram. So the UDP checksum offload was disabled by default. You can enable it by setting link0 flag with ifconfig(8). o bus_dma(9) clean up. It now correctly set number of required DMA segments/size and removed incorrect use of BUS_DMA_ALLOCNOW flag in static allocations done via bus_dmamem_alloc(9). o Implemented ALTQ(9) support. o Implemented Tx side bus_dmamap_load_mbuf_sg(9) which can remove several book keeping chores orginated from call-back mechanism. Therefore gem_txdma_callback() was removed and its functionality was reimplemented in gem_load_txmbuf(). o Don't set GEM_TD_START_OF_PACKET flag until all remaining mbuf chains are set. I think it was a long standing bug and it caused fluctuating interrupts/CPU usage patterns while netperf test is in progress. Previously it seems that we race with the device. Because I don't have a documentation for GEM I'm not sure this is correct but almost all other documentations I have stated this implications on setting SOP mark in descriptor ring(e.g. hme(4)). o Borrowed gem_defrag() from ath(4) which is supposed to be much faster than m_defrag(9) since it's not need to defrag all mbuf chains. o gem_load_txmbuf() was changed to allow passed mbuf chains to free. Caller of gem_load_txmbuf() correctly handles freed mbuf chains. o In gem_start_locked(), added checks for availability of Tx descriptors before trying to load DMA maps which could save CPU cycles when number of available descriptors are low. Also, simplyfy IFF_DRV_OACTIVE detection logic. o Removed hard-coded function names in CTR macros and replaced it with __func__. o Moved statistics counter register access to gem_tick() to reduce number of PCI bus accesses. There is no reason to update statistics counters in interrupt handler. o Removed unnecessary call of gem_start_locked() in gem_ioctl(). Reviewed by: grehan (initial version), marius (with improvements and suggestions) Tested by: grehan (ppc), marius(sparc64)
2007-06-04 06:01:04 +00:00
case SIOCSIFCAP:
GEM_LOCK(sc);
ifp->if_capenable = ifr->ifr_reqcap;
if ((ifp->if_capenable & IFCAP_TXCSUM) != 0)
ifp->if_hwassist = sc->sc_csum_features;
else
ifp->if_hwassist = 0;
GEM_UNLOCK(sc);
break;
default:
error = ether_ioctl(ifp, cmd, data);
break;
}
return (error);
}
static void
gem_setladrf(struct gem_softc *sc)
{
struct ifnet *ifp = sc->sc_ifp;
struct ifmultiaddr *inm;
int i;
uint32_t hash[16];
uint32_t crc, v;
GEM_LOCK_ASSERT(sc, MA_OWNED);
/*
* Turn off the RX MAC and the hash filter as required by the Sun GEM
* programming restrictions.
*/
v = sc->sc_mac_rxcfg & ~GEM_MAC_RX_HASH_FILTER;
GEM_BANK1_WRITE_4(sc, GEM_MAC_RX_CONFIG, v);
o The FreeBSD bus_dmamap_sync(9) supports ored together flags for quite some time now so collapse calls accordingly. o Given that gem_load_txmbuf() is allowed to fail resulting in a packet drop also for quite some time now implement the functionality of gem_txcksum() by means of m_pullup(9), which de-obfuscates the code and allows to always retrieve the correct length of the IP header. o Add missing BUS_DMASYNC_PREREAD when syncing the control DMA maps in gem_rint() and gem_start_locked(). o Correct some bus_barrier(9) calls to do a read/write barrier as we do a read after a write. Add some missing ones in gem_mii_readreg() and gem_mii_writereg(). o According to the Apple GMAC driver, the GEM ASIC specification and the OpenSolaris eri(7D) the TX FIFO threshold has to be set to 0x4ff for the Gigabit variants and 0x100 for the ERI in order do avoid TX underruns. o In gem_init_locked(): - be conservative and enable the RX and TX MACs, - don't clear GEM_LINK otherwise we don't ever mark the link as up again if gem_init_locked() is called from gem_watchdog(), - remove superfluous setting of sc_ifflags. o Don't bother to check whether the interface is running or whether its queue is empty before calling gem_start_locked() in gem_tint(), the former will check these anyway. o Call gem_start_locked() in gem_watchdog() in order to try to get some more packets going. o In gem_mii_writereg() after reseting the PCS restore its configuration. GMAC testing: grehan, marcel MFC after: 2 weeks
2008-06-22 13:54:51 +00:00
GEM_BANK1_BARRIER(sc, GEM_MAC_RX_CONFIG, 4,
BUS_SPACE_BARRIER_READ | BUS_SPACE_BARRIER_WRITE);
if (!GEM_BANK1_BITWAIT(sc, GEM_MAC_RX_CONFIG, GEM_MAC_RX_HASH_FILTER |
GEM_MAC_RX_ENABLE, 0))
device_printf(sc->sc_dev,
"cannot disable RX MAC or hash filter\n");
o Revert the part of if_gem.c rev. 1.35 which added a call to gem_stop() to gem_attach() as the former access softc members not yet initialized at that time and gem_reset() actually is enough to stop the chip. [1] o Revise the use of gem_bitwait(); add bus_barrier() calls before calling gem_bitwait() to ensure the respective bit has been written before we starting polling on it and poll for the right bits to change, f.e. even though we only reset RX we have to actually wait for both GEM_RESET_RX and GEM_RESET_TX to clear. Add some additional gem_bitwait() calls in places we've been missing them according to the GEM documentation. Along with this some excessive DELAYs, which probably only were added because of bugs in gem_bitwait() and its use in the first place, as well as as have of an gem_bitwait() reimplementation in gem_reset_tx() were removed. o Add gem_reset_rxdma() and use it to deal with GEM_MAC_RX_OVERFLOW errors more gracefully as unlike gem_init_locked() it resets the RX DMA engine only, causing no link loss and the FIFOs not to be cleared. Also use it deal with GEM_INTR_RX_TAG_ERR errors, with previously were unhandled. This was based on information obtained from the Linux GEM and OpenSolaris ERI drivers. o Turn on workarounds for silicon bugs in the Apple GMAC variants. This was based on information obtained from the Darwin GMAC and Linux GEM drivers. o Turn on "infinite" (i.e. maximum 31 * 64 bytes in length) DMA bursts. This greatly improves especially RX performance. o Optimize the RX path, this consists of: - kicking the receiver as soon as we've a spare descriptor in gem_rint() again instead of just once after all the ready ones have been handled; - kicking the receiver the right way, i.e. as outlined in the GEM documentation in batches of 4 and by pointing it to the descriptor after the last valid one; - calling gem_rint() before gem_tint() in gem_intr() as gem_tint() may take quite a while; - doubling the size of the RX ring to 256 descriptors. Overall the RX performance of a GEM in a 1GHz Sun Fire V210 was improved from ~100Mbit/s to ~850Mbit/s. o In gem_add_rxbuf() don't assign the newly allocated mbuf to rxs_mbuf before calling bus_dmamap_load_mbuf_sg(), if bus_dmamap_load_mbuf_sg() fails we'll free the newly allocated mbuf, unable to recycle the previous one but a NULL pointer dereference instead. o In gem_init_locked() honor the return value of gem_meminit(). o Simplify gem_ringsize() and dont' return garbage in the default case. Based on OpenBSD. o Don't turn on MAC control, MIF and PCS interrupts unless GEM_DEBUG is defined as we don't need/use these interrupts for operation. o In gem_start_locked() sync the DMA maps of the descriptor rings before every kick of the transmitter and not just once after enqueuing all packets as the NIC might instantly start transmitting after we kicked it the first time. o Keep state of the link state and use it to enable or disable the MAC in gem_mii_statchg() accordingly as well as to return early from gem_start_locked() in case the link is down. [3] o Initialize the maximum frame size to a sane value. o In gem_mii_statchg() enable carrier extension if appropriate. o Increment if_ierrors in case of an GEM_MAC_RX_OVERFLOW error and in gem_eint(). [3] o Handle IFF_ALLMULTI correctly; don't set it if we've turned promiscuous group mode on and don't clear the flag if we've disabled promiscuous group mode (these were mostly NOPs though). [2] o Let gem_eint() also report GEM_INTR_PERR errors. o Move setting sc_variant from gem_pci_probe() to gem_pci_attach() as device probe methods are not supposed to touch the softc. o Collapse sc_inited and sc_pci into bits for sc_flags. o Add CTASSERTs ensuring that GEM_NRXDESC and GEM_NTXDESC are set to legal values. o Correctly set up for 802.3x flow control, though #ifdef out the code that actually enables it as this needs more testing and mainly a proper framework to support it. o Correct and add some conversions from hard-coded functions names to __func__ which were borked or forgotten in if_gem.c rev. 1.42. o Use PCIR_BAR instead of a homegrown macro. o Replace sc_enaddr[6] with sc_enaddr[ETHER_ADDR_LEN]. o In gem_pci_attach() in case attaching fails release the resources in the opposite order they were allocated. o Make gem_reset() static to if_gem.c as it's not needed outside that module. o Remove the GEM_GIGABIT flag and the associated code; GEM_GIGABIT was never set and the associated code was in the wrong place. o Remove sc_mif_config; it was only used to cache the contents of the respective register within gem_attach(). o Remove the #ifdef'ed out NetBSD/OpenBSD code for establishing a suspend hook as it will never be used on FreeBSD. o Also probe Apple Intrepid 2 GMAC and Apple Shasta GMAC, add support for Apple K2 GMAC. Based on OpenBSD. o Add support for Sun GBE/P cards, or in other words actually add support for cards based on GEM to gem(4). This mainly consists of adding support for the TBI of these chips. Along with this the PHY selection code was rewritten to hardcode the PHY number for certain configurations as for example the PHY of the on-board ERI of Blade 1000 shows up twice causing no link as the second incarnation is isolated. These changes were ported from OpenBSD with some additional improvements and modulo some bugs. o Add code to if_gem_pci.c allowing to read the MAC-address from the VPD on systems without Open Firmware. This is an improved version of my variant of the respective code in if_hme_pci.c o Now that gem(4) is MI enable it for all archs. Pointed out by: yongari [1] Suggested by: rwatson [2], yongari [3] Tested on: i386 (GEM), powerpc (GMACs by marcel and yongari), sparc64 (ERI and GEM) Reviewed by: yongari Approved by: re (kensmith)
2007-09-26 21:14:18 +00:00
v &= ~(GEM_MAC_RX_PROMISCUOUS | GEM_MAC_RX_PROMISC_GRP);
if ((ifp->if_flags & IFF_PROMISC) != 0) {
v |= GEM_MAC_RX_PROMISCUOUS;
goto chipit;
}
if ((ifp->if_flags & IFF_ALLMULTI) != 0) {
v |= GEM_MAC_RX_PROMISC_GRP;
goto chipit;
}
/*
* Set up multicast address filter by passing all multicast
* addresses through a crc generator, and then using the high
* order 8 bits as an index into the 256 bit logical address
* filter. The high order 4 bits selects the word, while the
* other 4 bits select the bit within the word (where bit 0
* is the MSB).
*/
/* Clear the hash table. */
memset(hash, 0, sizeof(hash));
if_maddr_rlock(ifp);
TAILQ_FOREACH(inm, &ifp->if_multiaddrs, ifma_link) {
if (inm->ifma_addr->sa_family != AF_LINK)
continue;
crc = ether_crc32_le(LLADDR((struct sockaddr_dl *)
inm->ifma_addr), ETHER_ADDR_LEN);
/* We just want the 8 most significant bits. */
crc >>= 24;
/* Set the corresponding bit in the filter. */
hash[crc >> 4] |= 1 << (15 - (crc & 15));
}
if_maddr_runlock(ifp);
v |= GEM_MAC_RX_HASH_FILTER;
/* Now load the hash table into the chip (if we are using it). */
for (i = 0; i < 16; i++)
GEM_BANK1_WRITE_4(sc,
GEM_MAC_HASH0 + i * (GEM_MAC_HASH1 - GEM_MAC_HASH0),
hash[i]);
chipit:
sc->sc_mac_rxcfg = v;
GEM_BANK1_WRITE_4(sc, GEM_MAC_RX_CONFIG, v | GEM_MAC_RX_ENABLE);
}