freebsd-skq/sys/dev/sis/if_sis.c

2482 lines
59 KiB
C
Raw Normal View History

/*-
* Copyright (c) 2005 Poul-Henning Kamp <phk@FreeBSD.org>
* Copyright (c) 1997, 1998, 1999
* Bill Paul <wpaul@ctr.columbia.edu>. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* 3. All advertising materials mentioning features or use of this software
* must display the following acknowledgement:
* This product includes software developed by Bill Paul.
* 4. Neither the name of the author nor the names of any co-contributors
* may be used to endorse or promote products derived from this software
* without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY Bill Paul AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL Bill Paul OR THE VOICES IN HIS HEAD
* BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
* CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
* SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
* INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
* CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
* ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
* THE POSSIBILITY OF SUCH DAMAGE.
*/
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
/*
* SiS 900/SiS 7016 fast ethernet PCI NIC driver. Datasheets are
* available from http://www.sis.com.tw.
*
* This driver also supports the NatSemi DP83815. Datasheets are
* available from http://www.national.com.
*
* Written by Bill Paul <wpaul@ee.columbia.edu>
* Electrical Engineering Department
* Columbia University, New York City
*/
/*
* The SiS 900 is a fairly simple chip. It uses bus master DMA with
* simple TX and RX descriptors of 3 longwords in size. The receiver
* has a single perfect filter entry for the station address and a
* 128-bit multicast hash table. The SiS 900 has a built-in MII-based
* transceiver while the 7016 requires an external transceiver chip.
* Both chips offer the standard bit-bang MII interface as well as
* an enchanced PHY interface which simplifies accessing MII registers.
*
* The only downside to this chipset is that RX descriptors must be
* longword aligned.
*/
#ifdef HAVE_KERNEL_OPTION_HEADERS
#include "opt_device_polling.h"
#endif
#include <sys/param.h>
#include <sys/systm.h>
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
#include <sys/bus.h>
#include <sys/endian.h>
#include <sys/kernel.h>
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
#include <sys/lock.h>
#include <sys/malloc.h>
#include <sys/mbuf.h>
#include <sys/module.h>
#include <sys/socket.h>
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
#include <sys/sockio.h>
#include <sys/sysctl.h>
#include <net/if.h>
#include <net/if_arp.h>
#include <net/ethernet.h>
#include <net/if_dl.h>
#include <net/if_media.h>
2001-12-05 09:34:28 +00:00
#include <net/if_types.h>
#include <net/if_vlan_var.h>
#include <net/bpf.h>
#include <machine/bus.h>
#include <machine/resource.h>
#include <sys/bus.h>
#include <sys/rman.h>
#include <dev/mii/mii.h>
#include <dev/mii/miivar.h>
#include <dev/pci/pcireg.h>
#include <dev/pci/pcivar.h>
#define SIS_USEIOSPACE
#include <dev/sis/if_sisreg.h>
MODULE_DEPEND(sis, pci, 1, 1, 1);
MODULE_DEPEND(sis, ether, 1, 1, 1);
MODULE_DEPEND(sis, miibus, 1, 1, 1);
/* "device miibus" required. See GENERIC if you get errors here. */
#include "miibus_if.h"
#define SIS_LOCK(_sc) mtx_lock(&(_sc)->sis_mtx)
#define SIS_UNLOCK(_sc) mtx_unlock(&(_sc)->sis_mtx)
#define SIS_LOCK_ASSERT(_sc) mtx_assert(&(_sc)->sis_mtx, MA_OWNED)
/*
* register space access macros
*/
#define CSR_WRITE_4(sc, reg, val) bus_write_4(sc->sis_res[0], reg, val)
#define CSR_READ_4(sc, reg) bus_read_4(sc->sis_res[0], reg)
#define CSR_READ_2(sc, reg) bus_read_2(sc->sis_res[0], reg)
/*
* Various supported device vendors/types and their names.
*/
static struct sis_type sis_devs[] = {
{ SIS_VENDORID, SIS_DEVICEID_900, "SiS 900 10/100BaseTX" },
{ SIS_VENDORID, SIS_DEVICEID_7016, "SiS 7016 10/100BaseTX" },
{ NS_VENDORID, NS_DEVICEID_DP83815, "NatSemi DP8381[56] 10/100BaseTX" },
{ 0, 0, NULL }
};
static int sis_detach(device_t);
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
static __inline void sis_discard_rxbuf(struct sis_rxdesc *);
static int sis_dma_alloc(struct sis_softc *);
static void sis_dma_free(struct sis_softc *);
static int sis_dma_ring_alloc(struct sis_softc *, bus_size_t, bus_size_t,
bus_dma_tag_t *, uint8_t **, bus_dmamap_t *, bus_addr_t *, const char *);
static void sis_dmamap_cb(void *, bus_dma_segment_t *, int, int);
#ifndef __NO_STRICT_ALIGNMENT
static __inline void sis_fixup_rx(struct mbuf *);
#endif
static void sis_ifmedia_sts(struct ifnet *, struct ifmediareq *);
static int sis_ifmedia_upd(struct ifnet *);
static void sis_init(void *);
static void sis_initl(struct sis_softc *);
static void sis_intr(void *);
static int sis_ioctl(struct ifnet *, u_long, caddr_t);
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
static int sis_newbuf(struct sis_softc *, struct sis_rxdesc *);
static int sis_rxeof(struct sis_softc *);
static void sis_start(struct ifnet *);
static void sis_startl(struct ifnet *);
static void sis_stop(struct sis_softc *);
static void sis_add_sysctls(struct sis_softc *);
static void sis_watchdog(struct sis_softc *);
static struct resource_spec sis_res_spec[] = {
#ifdef SIS_USEIOSPACE
{ SYS_RES_IOPORT, SIS_PCI_LOIO, RF_ACTIVE},
#else
{ SYS_RES_MEMORY, SIS_PCI_LOMEM, RF_ACTIVE},
#endif
{ SYS_RES_IRQ, 0, RF_ACTIVE | RF_SHAREABLE},
{ -1, 0 }
};
#define SIS_SETBIT(sc, reg, x) \
CSR_WRITE_4(sc, reg, \
CSR_READ_4(sc, reg) | (x))
#define SIS_CLRBIT(sc, reg, x) \
CSR_WRITE_4(sc, reg, \
CSR_READ_4(sc, reg) & ~(x))
#define SIO_SET(x) \
CSR_WRITE_4(sc, SIS_EECTL, CSR_READ_4(sc, SIS_EECTL) | x)
#define SIO_CLR(x) \
CSR_WRITE_4(sc, SIS_EECTL, CSR_READ_4(sc, SIS_EECTL) & ~x)
/*
* Routine to reverse the bits in a word. Stolen almost
* verbatim from /usr/games/fortune.
*/
static uint16_t
sis_reverse(uint16_t n)
{
n = ((n >> 1) & 0x5555) | ((n << 1) & 0xaaaa);
n = ((n >> 2) & 0x3333) | ((n << 2) & 0xcccc);
n = ((n >> 4) & 0x0f0f) | ((n << 4) & 0xf0f0);
n = ((n >> 8) & 0x00ff) | ((n << 8) & 0xff00);
2010-09-01 18:39:35 +00:00
return (n);
}
static void
sis_delay(struct sis_softc *sc)
{
int idx;
for (idx = (300 / 33) + 1; idx > 0; idx--)
CSR_READ_4(sc, SIS_CSR);
}
static void
sis_eeprom_idle(struct sis_softc *sc)
{
int i;
SIO_SET(SIS_EECTL_CSEL);
sis_delay(sc);
SIO_SET(SIS_EECTL_CLK);
sis_delay(sc);
for (i = 0; i < 25; i++) {
SIO_CLR(SIS_EECTL_CLK);
sis_delay(sc);
SIO_SET(SIS_EECTL_CLK);
sis_delay(sc);
}
SIO_CLR(SIS_EECTL_CLK);
sis_delay(sc);
SIO_CLR(SIS_EECTL_CSEL);
sis_delay(sc);
CSR_WRITE_4(sc, SIS_EECTL, 0x00000000);
}
/*
* Send a read command and address to the EEPROM, check for ACK.
*/
static void
sis_eeprom_putbyte(struct sis_softc *sc, int addr)
{
int d, i;
d = addr | SIS_EECMD_READ;
/*
* Feed in each bit and stobe the clock.
*/
for (i = 0x400; i; i >>= 1) {
if (d & i) {
SIO_SET(SIS_EECTL_DIN);
} else {
SIO_CLR(SIS_EECTL_DIN);
}
sis_delay(sc);
SIO_SET(SIS_EECTL_CLK);
sis_delay(sc);
SIO_CLR(SIS_EECTL_CLK);
sis_delay(sc);
}
}
/*
* Read a word of data stored in the EEPROM at address 'addr.'
*/
static void
sis_eeprom_getword(struct sis_softc *sc, int addr, uint16_t *dest)
{
int i;
2010-09-01 19:53:15 +00:00
uint16_t word = 0;
/* Force EEPROM to idle state. */
sis_eeprom_idle(sc);
/* Enter EEPROM access mode. */
sis_delay(sc);
SIO_CLR(SIS_EECTL_CLK);
sis_delay(sc);
SIO_SET(SIS_EECTL_CSEL);
sis_delay(sc);
/*
* Send address of word we want to read.
*/
sis_eeprom_putbyte(sc, addr);
/*
* Start reading bits from EEPROM.
*/
for (i = 0x8000; i; i >>= 1) {
SIO_SET(SIS_EECTL_CLK);
sis_delay(sc);
if (CSR_READ_4(sc, SIS_EECTL) & SIS_EECTL_DOUT)
word |= i;
sis_delay(sc);
SIO_CLR(SIS_EECTL_CLK);
sis_delay(sc);
}
/* Turn off EEPROM access mode. */
sis_eeprom_idle(sc);
*dest = word;
}
/*
* Read a sequence of words from the EEPROM.
*/
static void
sis_read_eeprom(struct sis_softc *sc, caddr_t dest, int off, int cnt, int swap)
{
int i;
2010-09-01 19:53:15 +00:00
uint16_t word = 0, *ptr;
for (i = 0; i < cnt; i++) {
sis_eeprom_getword(sc, off + i, &word);
2010-09-01 19:53:15 +00:00
ptr = (uint16_t *)(dest + (i * 2));
if (swap)
*ptr = ntohs(word);
else
*ptr = word;
}
}
#if defined(__i386__) || defined(__amd64__)
static device_t
sis_find_bridge(device_t dev)
{
devclass_t pci_devclass;
device_t *pci_devices;
int pci_count = 0;
device_t *pci_children;
int pci_childcount = 0;
device_t *busp, *childp;
device_t child = NULL;
int i, j;
if ((pci_devclass = devclass_find("pci")) == NULL)
2010-09-01 18:39:35 +00:00
return (NULL);
devclass_get_devices(pci_devclass, &pci_devices, &pci_count);
for (i = 0, busp = pci_devices; i < pci_count; i++, busp++) {
if (device_get_children(*busp, &pci_children, &pci_childcount))
continue;
for (j = 0, childp = pci_children;
j < pci_childcount; j++, childp++) {
if (pci_get_vendor(*childp) == SIS_VENDORID &&
pci_get_device(*childp) == 0x0008) {
child = *childp;
free(pci_children, M_TEMP);
goto done;
}
}
free(pci_children, M_TEMP);
}
done:
free(pci_devices, M_TEMP);
2010-09-01 18:39:35 +00:00
return (child);
}
static void
sis_read_cmos(struct sis_softc *sc, device_t dev, caddr_t dest, int off, int cnt)
{
device_t bridge;
2010-09-01 19:53:15 +00:00
uint8_t reg;
int i;
bus_space_tag_t btag;
bridge = sis_find_bridge(dev);
if (bridge == NULL)
return;
reg = pci_read_config(bridge, 0x48, 1);
pci_write_config(bridge, 0x48, reg|0x40, 1);
/* XXX */
#if defined(__i386__)
btag = I386_BUS_SPACE_IO;
#elif defined(__amd64__)
btag = AMD64_BUS_SPACE_IO;
#endif
for (i = 0; i < cnt; i++) {
bus_space_write_1(btag, 0x0, 0x70, i + off);
*(dest + i) = bus_space_read_1(btag, 0x0, 0x71);
}
pci_write_config(bridge, 0x48, reg & ~0x40, 1);
}
static void
sis_read_mac(struct sis_softc *sc, device_t dev, caddr_t dest)
{
2010-09-01 19:53:15 +00:00
uint32_t filtsave, csrsave;
filtsave = CSR_READ_4(sc, SIS_RXFILT_CTL);
csrsave = CSR_READ_4(sc, SIS_CSR);
CSR_WRITE_4(sc, SIS_CSR, SIS_CSR_RELOAD | filtsave);
CSR_WRITE_4(sc, SIS_CSR, 0);
CSR_WRITE_4(sc, SIS_RXFILT_CTL, filtsave & ~SIS_RXFILTCTL_ENABLE);
CSR_WRITE_4(sc, SIS_RXFILT_CTL, SIS_FILTADDR_PAR0);
2010-09-01 19:53:15 +00:00
((uint16_t *)dest)[0] = CSR_READ_2(sc, SIS_RXFILT_DATA);
CSR_WRITE_4(sc, SIS_RXFILT_CTL,SIS_FILTADDR_PAR1);
2010-09-01 19:53:15 +00:00
((uint16_t *)dest)[1] = CSR_READ_2(sc, SIS_RXFILT_DATA);
CSR_WRITE_4(sc, SIS_RXFILT_CTL, SIS_FILTADDR_PAR2);
2010-09-01 19:53:15 +00:00
((uint16_t *)dest)[2] = CSR_READ_2(sc, SIS_RXFILT_DATA);
CSR_WRITE_4(sc, SIS_RXFILT_CTL, filtsave);
CSR_WRITE_4(sc, SIS_CSR, csrsave);
}
#endif
/*
* Sync the PHYs by setting data bit and strobing the clock 32 times.
*/
static void
sis_mii_sync(struct sis_softc *sc)
{
int i;
SIO_SET(SIS_MII_DIR|SIS_MII_DATA);
for (i = 0; i < 32; i++) {
SIO_SET(SIS_MII_CLK);
DELAY(1);
SIO_CLR(SIS_MII_CLK);
DELAY(1);
}
}
/*
* Clock a series of bits through the MII.
*/
static void
sis_mii_send(struct sis_softc *sc, uint32_t bits, int cnt)
{
int i;
SIO_CLR(SIS_MII_CLK);
for (i = (0x1 << (cnt - 1)); i; i >>= 1) {
if (bits & i) {
SIO_SET(SIS_MII_DATA);
} else {
SIO_CLR(SIS_MII_DATA);
}
DELAY(1);
SIO_CLR(SIS_MII_CLK);
DELAY(1);
SIO_SET(SIS_MII_CLK);
}
}
/*
* Read an PHY register through the MII.
*/
static int
sis_mii_readreg(struct sis_softc *sc, struct sis_mii_frame *frame)
{
int i, ack;
/*
* Set up frame for RX.
*/
frame->mii_stdelim = SIS_MII_STARTDELIM;
frame->mii_opcode = SIS_MII_READOP;
frame->mii_turnaround = 0;
frame->mii_data = 0;
/*
* Turn on data xmit.
*/
SIO_SET(SIS_MII_DIR);
sis_mii_sync(sc);
/*
* Send command/address info.
*/
sis_mii_send(sc, frame->mii_stdelim, 2);
sis_mii_send(sc, frame->mii_opcode, 2);
sis_mii_send(sc, frame->mii_phyaddr, 5);
sis_mii_send(sc, frame->mii_regaddr, 5);
/* Idle bit */
SIO_CLR((SIS_MII_CLK|SIS_MII_DATA));
DELAY(1);
SIO_SET(SIS_MII_CLK);
DELAY(1);
/* Turn off xmit. */
SIO_CLR(SIS_MII_DIR);
/* Check for ack */
SIO_CLR(SIS_MII_CLK);
DELAY(1);
ack = CSR_READ_4(sc, SIS_EECTL) & SIS_MII_DATA;
SIO_SET(SIS_MII_CLK);
DELAY(1);
/*
* Now try reading data bits. If the ack failed, we still
* need to clock through 16 cycles to keep the PHY(s) in sync.
*/
if (ack) {
2010-09-01 18:39:35 +00:00
for (i = 0; i < 16; i++) {
SIO_CLR(SIS_MII_CLK);
DELAY(1);
SIO_SET(SIS_MII_CLK);
DELAY(1);
}
goto fail;
}
for (i = 0x8000; i; i >>= 1) {
SIO_CLR(SIS_MII_CLK);
DELAY(1);
if (!ack) {
if (CSR_READ_4(sc, SIS_EECTL) & SIS_MII_DATA)
frame->mii_data |= i;
DELAY(1);
}
SIO_SET(SIS_MII_CLK);
DELAY(1);
}
fail:
SIO_CLR(SIS_MII_CLK);
DELAY(1);
SIO_SET(SIS_MII_CLK);
DELAY(1);
if (ack)
2010-09-01 18:39:35 +00:00
return (1);
return (0);
}
/*
* Write to a PHY register through the MII.
*/
static int
sis_mii_writereg(struct sis_softc *sc, struct sis_mii_frame *frame)
{
/*
* Set up frame for TX.
*/
frame->mii_stdelim = SIS_MII_STARTDELIM;
frame->mii_opcode = SIS_MII_WRITEOP;
frame->mii_turnaround = SIS_MII_TURNAROUND;
/*
* Turn on data output.
*/
SIO_SET(SIS_MII_DIR);
sis_mii_sync(sc);
sis_mii_send(sc, frame->mii_stdelim, 2);
sis_mii_send(sc, frame->mii_opcode, 2);
sis_mii_send(sc, frame->mii_phyaddr, 5);
sis_mii_send(sc, frame->mii_regaddr, 5);
sis_mii_send(sc, frame->mii_turnaround, 2);
sis_mii_send(sc, frame->mii_data, 16);
/* Idle bit. */
SIO_SET(SIS_MII_CLK);
DELAY(1);
SIO_CLR(SIS_MII_CLK);
DELAY(1);
/*
* Turn off xmit.
*/
SIO_CLR(SIS_MII_DIR);
2010-09-01 18:39:35 +00:00
return (0);
}
static int
sis_miibus_readreg(device_t dev, int phy, int reg)
{
struct sis_softc *sc;
struct sis_mii_frame frame;
sc = device_get_softc(dev);
if (sc->sis_type == SIS_TYPE_83815) {
if (phy != 0)
2010-09-01 18:39:35 +00:00
return (0);
/*
* The NatSemi chip can take a while after
* a reset to come ready, during which the BMSR
* returns a value of 0. This is *never* supposed
* to happen: some of the BMSR bits are meant to
* be hardwired in the on position, and this can
* confuse the miibus code a bit during the probe
* and attach phase. So we make an effort to check
* for this condition and wait for it to clear.
*/
if (!CSR_READ_4(sc, NS_BMSR))
DELAY(1000);
return CSR_READ_4(sc, NS_BMCR + (reg * 4));
}
/*
* Chipsets < SIS_635 seem not to be able to read/write
* through mdio. Use the enhanced PHY access register
* again for them.
*/
if (sc->sis_type == SIS_TYPE_900 &&
sc->sis_rev < SIS_REV_635) {
int i, val = 0;
if (phy != 0)
2010-09-01 18:39:35 +00:00
return (0);
CSR_WRITE_4(sc, SIS_PHYCTL,
(phy << 11) | (reg << 6) | SIS_PHYOP_READ);
SIS_SETBIT(sc, SIS_PHYCTL, SIS_PHYCTL_ACCESS);
for (i = 0; i < SIS_TIMEOUT; i++) {
if (!(CSR_READ_4(sc, SIS_PHYCTL) & SIS_PHYCTL_ACCESS))
break;
}
if (i == SIS_TIMEOUT) {
device_printf(sc->sis_dev, "PHY failed to come ready\n");
2010-09-01 18:39:35 +00:00
return (0);
}
val = (CSR_READ_4(sc, SIS_PHYCTL) >> 16) & 0xFFFF;
if (val == 0xFFFF)
2010-09-01 18:39:35 +00:00
return (0);
2010-09-01 18:39:35 +00:00
return (val);
} else {
bzero((char *)&frame, sizeof(frame));
frame.mii_phyaddr = phy;
frame.mii_regaddr = reg;
sis_mii_readreg(sc, &frame);
2010-09-01 18:39:35 +00:00
return (frame.mii_data);
}
}
static int
sis_miibus_writereg(device_t dev, int phy, int reg, int data)
{
struct sis_softc *sc;
struct sis_mii_frame frame;
sc = device_get_softc(dev);
if (sc->sis_type == SIS_TYPE_83815) {
if (phy != 0)
2010-09-01 18:39:35 +00:00
return (0);
CSR_WRITE_4(sc, NS_BMCR + (reg * 4), data);
2010-09-01 18:39:35 +00:00
return (0);
}
/*
* Chipsets < SIS_635 seem not to be able to read/write
* through mdio. Use the enhanced PHY access register
* again for them.
*/
if (sc->sis_type == SIS_TYPE_900 &&
sc->sis_rev < SIS_REV_635) {
int i;
if (phy != 0)
2010-09-01 18:39:35 +00:00
return (0);
CSR_WRITE_4(sc, SIS_PHYCTL, (data << 16) | (phy << 11) |
(reg << 6) | SIS_PHYOP_WRITE);
SIS_SETBIT(sc, SIS_PHYCTL, SIS_PHYCTL_ACCESS);
for (i = 0; i < SIS_TIMEOUT; i++) {
if (!(CSR_READ_4(sc, SIS_PHYCTL) & SIS_PHYCTL_ACCESS))
break;
}
if (i == SIS_TIMEOUT)
device_printf(sc->sis_dev, "PHY failed to come ready\n");
} else {
bzero((char *)&frame, sizeof(frame));
frame.mii_phyaddr = phy;
frame.mii_regaddr = reg;
frame.mii_data = data;
sis_mii_writereg(sc, &frame);
}
2010-09-01 18:39:35 +00:00
return (0);
}
static void
sis_miibus_statchg(device_t dev)
{
struct sis_softc *sc;
struct mii_data *mii;
struct ifnet *ifp;
uint32_t reg;
sc = device_get_softc(dev);
SIS_LOCK_ASSERT(sc);
mii = device_get_softc(sc->sis_miibus);
ifp = sc->sis_ifp;
if (mii == NULL || ifp == NULL ||
(ifp->if_drv_flags & IFF_DRV_RUNNING) == 0)
return;
sc->sis_flags &= ~SIS_FLAG_LINK;
if ((mii->mii_media_status & (IFM_ACTIVE | IFM_AVALID)) ==
(IFM_ACTIVE | IFM_AVALID)) {
switch (IFM_SUBTYPE(mii->mii_media_active)) {
case IFM_10_T:
CSR_WRITE_4(sc, SIS_TX_CFG, SIS_TXCFG_10);
sc->sis_flags |= SIS_FLAG_LINK;
break;
case IFM_100_TX:
CSR_WRITE_4(sc, SIS_TX_CFG, SIS_TXCFG_100);
sc->sis_flags |= SIS_FLAG_LINK;
break;
default:
break;
}
}
if ((sc->sis_flags & SIS_FLAG_LINK) == 0) {
/*
* Stopping MACs seem to reset SIS_TX_LISTPTR and
* SIS_RX_LISTPTR which in turn requires resetting
* TX/RX buffers. So just don't do anything for
* lost link.
*/
return;
}
/* Set full/half duplex mode. */
if ((IFM_OPTIONS(mii->mii_media_active) & IFM_FDX) != 0) {
SIS_SETBIT(sc, SIS_TX_CFG,
(SIS_TXCFG_IGN_HBEAT | SIS_TXCFG_IGN_CARR));
SIS_SETBIT(sc, SIS_RX_CFG, SIS_RXCFG_RX_TXPKTS);
} else {
SIS_CLRBIT(sc, SIS_TX_CFG,
(SIS_TXCFG_IGN_HBEAT | SIS_TXCFG_IGN_CARR));
SIS_CLRBIT(sc, SIS_RX_CFG, SIS_RXCFG_RX_TXPKTS);
}
if (sc->sis_type == SIS_TYPE_83816) {
/*
* MPII03.D: Half Duplex Excessive Collisions.
* Also page 49 in 83816 manual
*/
SIS_SETBIT(sc, SIS_TX_CFG, SIS_TXCFG_MPII03D);
}
if (sc->sis_type == SIS_TYPE_83815 && sc->sis_srr < NS_SRR_16A &&
IFM_SUBTYPE(mii->mii_media_active) == IFM_100_TX) {
/*
* Short Cable Receive Errors (MP21.E)
*/
CSR_WRITE_4(sc, NS_PHY_PAGE, 0x0001);
reg = CSR_READ_4(sc, NS_PHY_DSPCFG) & 0xfff;
CSR_WRITE_4(sc, NS_PHY_DSPCFG, reg | 0x1000);
DELAY(100);
reg = CSR_READ_4(sc, NS_PHY_TDATA) & 0xff;
if ((reg & 0x0080) == 0 || (reg > 0xd8 && reg <= 0xff)) {
device_printf(sc->sis_dev,
"Applying short cable fix (reg=%x)\n", reg);
CSR_WRITE_4(sc, NS_PHY_TDATA, 0x00e8);
SIS_SETBIT(sc, NS_PHY_DSPCFG, 0x20);
}
CSR_WRITE_4(sc, NS_PHY_PAGE, 0);
}
/* Enable TX/RX MACs. */
SIS_CLRBIT(sc, SIS_CSR, SIS_CSR_TX_DISABLE | SIS_CSR_RX_DISABLE);
SIS_SETBIT(sc, SIS_CSR, SIS_CSR_TX_ENABLE | SIS_CSR_RX_ENABLE);
}
static uint32_t
sis_mchash(struct sis_softc *sc, const uint8_t *addr)
{
uint32_t crc;
/* Compute CRC for the address value. */
crc = ether_crc32_be(addr, ETHER_ADDR_LEN);
/*
* return the filter bit position
*
* The NatSemi chip has a 512-bit filter, which is
* different than the SiS, so we special-case it.
*/
if (sc->sis_type == SIS_TYPE_83815)
return (crc >> 23);
else if (sc->sis_rev >= SIS_REV_635 ||
sc->sis_rev == SIS_REV_900B)
return (crc >> 24);
else
return (crc >> 25);
}
static void
sis_setmulti_ns(struct sis_softc *sc)
{
struct ifnet *ifp;
struct ifmultiaddr *ifma;
2010-09-01 19:53:15 +00:00
uint32_t h = 0, i, filtsave;
int bit, index;
ifp = sc->sis_ifp;
if (ifp->if_flags & IFF_ALLMULTI || ifp->if_flags & IFF_PROMISC) {
SIS_CLRBIT(sc, SIS_RXFILT_CTL, NS_RXFILTCTL_MCHASH);
SIS_SETBIT(sc, SIS_RXFILT_CTL, SIS_RXFILTCTL_ALLMULTI);
return;
}
/*
* We have to explicitly enable the multicast hash table
* on the NatSemi chip if we want to use it, which we do.
*/
SIS_SETBIT(sc, SIS_RXFILT_CTL, NS_RXFILTCTL_MCHASH);
SIS_CLRBIT(sc, SIS_RXFILT_CTL, SIS_RXFILTCTL_ALLMULTI);
filtsave = CSR_READ_4(sc, SIS_RXFILT_CTL);
/* first, zot all the existing hash bits */
for (i = 0; i < 32; i++) {
CSR_WRITE_4(sc, SIS_RXFILT_CTL, NS_FILTADDR_FMEM_LO + (i*2));
CSR_WRITE_4(sc, SIS_RXFILT_DATA, 0);
}
if_maddr_rlock(ifp);
TAILQ_FOREACH(ifma, &ifp->if_multiaddrs, ifma_link) {
if (ifma->ifma_addr->sa_family != AF_LINK)
continue;
h = sis_mchash(sc,
LLADDR((struct sockaddr_dl *)ifma->ifma_addr));
index = h >> 3;
bit = h & 0x1F;
CSR_WRITE_4(sc, SIS_RXFILT_CTL, NS_FILTADDR_FMEM_LO + index);
if (bit > 0xF)
bit -= 0x10;
SIS_SETBIT(sc, SIS_RXFILT_DATA, (1 << bit));
}
if_maddr_runlock(ifp);
CSR_WRITE_4(sc, SIS_RXFILT_CTL, filtsave);
}
static void
sis_setmulti_sis(struct sis_softc *sc)
{
struct ifnet *ifp;
struct ifmultiaddr *ifma;
2010-09-01 19:53:15 +00:00
uint32_t h, i, n, ctl;
uint16_t hashes[16];
ifp = sc->sis_ifp;
/* hash table size */
if (sc->sis_rev >= SIS_REV_635 ||
sc->sis_rev == SIS_REV_900B)
n = 16;
else
n = 8;
ctl = CSR_READ_4(sc, SIS_RXFILT_CTL) & SIS_RXFILTCTL_ENABLE;
if (ifp->if_flags & IFF_BROADCAST)
ctl |= SIS_RXFILTCTL_BROAD;
if (ifp->if_flags & IFF_ALLMULTI || ifp->if_flags & IFF_PROMISC) {
ctl |= SIS_RXFILTCTL_ALLMULTI;
if (ifp->if_flags & IFF_PROMISC)
ctl |= SIS_RXFILTCTL_BROAD|SIS_RXFILTCTL_ALLPHYS;
for (i = 0; i < n; i++)
hashes[i] = ~0;
} else {
for (i = 0; i < n; i++)
hashes[i] = 0;
i = 0;
if_maddr_rlock(ifp);
TAILQ_FOREACH(ifma, &ifp->if_multiaddrs, ifma_link) {
if (ifma->ifma_addr->sa_family != AF_LINK)
continue;
h = sis_mchash(sc,
LLADDR((struct sockaddr_dl *)ifma->ifma_addr));
hashes[h >> 4] |= 1 << (h & 0xf);
i++;
}
if_maddr_runlock(ifp);
if (i > n) {
ctl |= SIS_RXFILTCTL_ALLMULTI;
for (i = 0; i < n; i++)
hashes[i] = ~0;
}
}
for (i = 0; i < n; i++) {
CSR_WRITE_4(sc, SIS_RXFILT_CTL, (4 + i) << 16);
CSR_WRITE_4(sc, SIS_RXFILT_DATA, hashes[i]);
}
CSR_WRITE_4(sc, SIS_RXFILT_CTL, ctl);
}
static void
sis_reset(struct sis_softc *sc)
{
int i;
SIS_SETBIT(sc, SIS_CSR, SIS_CSR_RESET);
for (i = 0; i < SIS_TIMEOUT; i++) {
if (!(CSR_READ_4(sc, SIS_CSR) & SIS_CSR_RESET))
break;
}
if (i == SIS_TIMEOUT)
device_printf(sc->sis_dev, "reset never completed\n");
/* Wait a little while for the chip to get its brains in order. */
DELAY(1000);
/*
* If this is a NetSemi chip, make sure to clear
* PME mode.
*/
if (sc->sis_type == SIS_TYPE_83815) {
CSR_WRITE_4(sc, NS_CLKRUN, NS_CLKRUN_PMESTS);
CSR_WRITE_4(sc, NS_CLKRUN, 0);
}
}
/*
* Probe for an SiS chip. Check the PCI vendor and device
* IDs against our list and return a device name if we find a match.
*/
static int
sis_probe(device_t dev)
{
struct sis_type *t;
t = sis_devs;
2010-09-01 18:39:35 +00:00
while (t->sis_name != NULL) {
if ((pci_get_vendor(dev) == t->sis_vid) &&
(pci_get_device(dev) == t->sis_did)) {
device_set_desc(dev, t->sis_name);
2005-02-24 21:32:56 +00:00
return (BUS_PROBE_DEFAULT);
}
t++;
}
2010-09-01 18:39:35 +00:00
return (ENXIO);
}
/*
* Attach the interface. Allocate softc structures, do ifmedia
* setup and ethernet/BPF attach.
*/
static int
sis_attach(device_t dev)
{
u_char eaddr[ETHER_ADDR_LEN];
struct sis_softc *sc;
struct ifnet *ifp;
int error = 0, waittime = 0;
waittime = 0;
sc = device_get_softc(dev);
sc->sis_dev = dev;
mtx_init(&sc->sis_mtx, device_get_nameunit(dev), MTX_NETWORK_LOCK,
2005-01-07 00:02:11 +00:00
MTX_DEF);
callout_init_mtx(&sc->sis_stat_ch, &sc->sis_mtx, 0);
if (pci_get_device(dev) == SIS_DEVICEID_900)
sc->sis_type = SIS_TYPE_900;
if (pci_get_device(dev) == SIS_DEVICEID_7016)
sc->sis_type = SIS_TYPE_7016;
if (pci_get_vendor(dev) == NS_VENDORID)
sc->sis_type = SIS_TYPE_83815;
sc->sis_rev = pci_read_config(dev, PCIR_REVID, 1);
/*
* Map control/status registers.
*/
pci_enable_busmaster(dev);
error = bus_alloc_resources(dev, sis_res_spec, sc->sis_res);
if (error) {
device_printf(dev, "couldn't allocate resources\n");
goto fail;
}
/* Reset the adapter. */
sis_reset(sc);
if (sc->sis_type == SIS_TYPE_900 &&
(sc->sis_rev == SIS_REV_635 ||
sc->sis_rev == SIS_REV_900B)) {
SIO_SET(SIS_CFG_RND_CNT);
SIO_SET(SIS_CFG_PERR_DETECT);
}
/*
* Get station address from the EEPROM.
*/
switch (pci_get_vendor(dev)) {
case NS_VENDORID:
sc->sis_srr = CSR_READ_4(sc, NS_SRR);
/* We can't update the device description, so spew */
if (sc->sis_srr == NS_SRR_15C)
device_printf(dev, "Silicon Revision: DP83815C\n");
else if (sc->sis_srr == NS_SRR_15D)
device_printf(dev, "Silicon Revision: DP83815D\n");
else if (sc->sis_srr == NS_SRR_16A)
device_printf(dev, "Silicon Revision: DP83816A\n");
else
device_printf(dev, "Silicon Revision %x\n", sc->sis_srr);
/*
* Reading the MAC address out of the EEPROM on
* the NatSemi chip takes a bit more work than
* you'd expect. The address spans 4 16-bit words,
* with the first word containing only a single bit.
* You have to shift everything over one bit to
* get it aligned properly. Also, the bits are
* stored backwards (the LSB is really the MSB,
* and so on) so you have to reverse them in order
* to get the MAC address into the form we want.
* Why? Who the hell knows.
*/
{
2010-09-01 19:53:15 +00:00
uint16_t tmp[4];
sis_read_eeprom(sc, (caddr_t)&tmp,
NS_EE_NODEADDR, 4, 0);
/* Shift everything over one bit. */
tmp[3] = tmp[3] >> 1;
tmp[3] |= tmp[2] << 15;
tmp[2] = tmp[2] >> 1;
tmp[2] |= tmp[1] << 15;
tmp[1] = tmp[1] >> 1;
tmp[1] |= tmp[0] << 15;
/* Now reverse all the bits. */
tmp[3] = sis_reverse(tmp[3]);
tmp[2] = sis_reverse(tmp[2]);
tmp[1] = sis_reverse(tmp[1]);
eaddr[0] = (tmp[1] >> 0) & 0xFF;
eaddr[1] = (tmp[1] >> 8) & 0xFF;
eaddr[2] = (tmp[2] >> 0) & 0xFF;
eaddr[3] = (tmp[2] >> 8) & 0xFF;
eaddr[4] = (tmp[3] >> 0) & 0xFF;
eaddr[5] = (tmp[3] >> 8) & 0xFF;
}
break;
case SIS_VENDORID:
default:
#if defined(__i386__) || defined(__amd64__)
/*
* If this is a SiS 630E chipset with an embedded
* SiS 900 controller, we have to read the MAC address
* from the APC CMOS RAM. Our method for doing this
* is very ugly since we have to reach out and grab
* ahold of hardware for which we cannot properly
* allocate resources. This code is only compiled on
* the i386 architecture since the SiS 630E chipset
* is for x86 motherboards only. Note that there are
* a lot of magic numbers in this hack. These are
* taken from SiS's Linux driver. I'd like to replace
* them with proper symbolic definitions, but that
* requires some datasheets that I don't have access
* to at the moment.
*/
if (sc->sis_rev == SIS_REV_630S ||
sc->sis_rev == SIS_REV_630E ||
sc->sis_rev == SIS_REV_630EA1)
sis_read_cmos(sc, dev, (caddr_t)&eaddr, 0x9, 6);
else if (sc->sis_rev == SIS_REV_635 ||
sc->sis_rev == SIS_REV_630ET)
sis_read_mac(sc, dev, (caddr_t)&eaddr);
else if (sc->sis_rev == SIS_REV_96x) {
/* Allow to read EEPROM from LAN. It is shared
* between a 1394 controller and the NIC and each
* time we access it, we need to set SIS_EECMD_REQ.
*/
SIO_SET(SIS_EECMD_REQ);
for (waittime = 0; waittime < SIS_TIMEOUT;
waittime++) {
/* Force EEPROM to idle state. */
sis_eeprom_idle(sc);
if (CSR_READ_4(sc, SIS_EECTL) & SIS_EECMD_GNT) {
sis_read_eeprom(sc, (caddr_t)&eaddr,
SIS_EE_NODEADDR, 3, 0);
break;
}
DELAY(1);
}
/*
* Set SIS_EECTL_CLK to high, so a other master
* can operate on the i2c bus.
*/
SIO_SET(SIS_EECTL_CLK);
/* Refuse EEPROM access by LAN */
SIO_SET(SIS_EECMD_DONE);
} else
#endif
sis_read_eeprom(sc, (caddr_t)&eaddr,
SIS_EE_NODEADDR, 3, 0);
break;
}
sis_add_sysctls(sc);
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
/* Allocate DMA'able memory. */
if ((error = sis_dma_alloc(sc)) != 0)
Clean up locking and resource management for pci/if_* - Remove locking of the softc in the attach method, instead depending on bus_setup_intr being at the end of attach (delaying interrupt enable until after ether_ifattach is called) - Call *_detach directly in the error case of attach, depending on checking in detach to only free resources that were allocated. This puts all resource freeing in one place, avoiding thinkos that lead to memory leaks. - Add bus_child_present check to calls to *_stop in the detach method to be sure hw is present before touching its registers. - Remove bzero softc calls since device_t should do this for us. - dc: move interrupt allocation back where it was before. It was unnecessary to move it. This reverts part of 1.88 - rl: move irq allocation before ether_ifattach. Problems might have been caused by allocating the irq after enabling interrupts on the card. - rl: call rl_stop before ether_ifdetach - sf: call sf_stop before ether_ifdetach - sis: add missed free of sis_tag - sis: check errors from tag creation - sis: move dmamem_alloc and dmamap_load to happen at same time as tag creation - sk: remove duplicate initialization of sk_dev - ste: add missed bus_generic_detach - ti: call ti_stop before ether_ifdetach - ti: add missed error setting in ti_rdata alloc failure - vr: add missed error setting in I/O, memory mapping cases - xl: add missed error setting in I/O, memory mapping cases - xl: remove multi-level goto on attach failure - xl: move dmamem_alloc and dmamap_load to happen at same time as tag creation - Calls to free(9) are unconditional because it is valid to call free with a null pointer. Reviewed by: imp, mdodd
2003-03-31 17:29:43 +00:00
goto fail;
ifp = sc->sis_ifp = if_alloc(IFT_ETHER);
if (ifp == NULL) {
device_printf(dev, "can not if_alloc()\n");
error = ENOSPC;
goto fail;
}
ifp->if_softc = sc;
if_initname(ifp, device_get_name(dev), device_get_unit(dev));
ifp->if_mtu = ETHERMTU;
ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
ifp->if_ioctl = sis_ioctl;
ifp->if_start = sis_start;
ifp->if_init = sis_init;
IFQ_SET_MAXLEN(&ifp->if_snd, SIS_TX_LIST_CNT - 1);
ifp->if_snd.ifq_drv_maxlen = SIS_TX_LIST_CNT - 1;
IFQ_SET_READY(&ifp->if_snd);
/*
* Do MII setup.
*/
if (mii_phy_probe(dev, &sc->sis_miibus,
sis_ifmedia_upd, sis_ifmedia_sts)) {
device_printf(dev, "MII without any PHY!\n");
error = ENXIO;
goto fail;
}
/*
* Call MI attach routine.
*/
ether_ifattach(ifp, eaddr);
2001-12-05 09:34:28 +00:00
/*
* Tell the upper layer(s) we support long frames.
*/
ifp->if_data.ifi_hdrlen = sizeof(struct ether_vlan_header);
ifp->if_capabilities |= IFCAP_VLAN_MTU;
Big polling(4) cleanup. o Axe poll in trap. o Axe IFF_POLLING flag from if_flags. o Rework revision 1.21 (Giant removal), in such a way that poll_mtx is not dropped during call to polling handler. This fixes problem with idle polling. o Make registration and deregistration from polling in a functional way, insted of next tick/interrupt. o Obsolete kern.polling.enable. Polling is turned on/off with ifconfig. Detailed kern_poll.c changes: - Remove polling handler flags, introduced in 1.21. The are not needed now. - Forget and do not check if_flags, if_capenable and if_drv_flags. - Call all registered polling handlers unconditionally. - Do not drop poll_mtx, when entering polling handlers. - In ether_poll() NET_LOCK_GIANT prior to locking poll_mtx. - In netisr_poll() axe the block, where polling code asks drivers to unregister. - In netisr_poll() and ether_poll() do polling always, if any handlers are present. - In ether_poll_[de]register() remove a lot of error hiding code. Assert that arguments are correct, instead. - In ether_poll_[de]register() use standard return values in case of error or success. - Introduce poll_switch() that is a sysctl handler for kern.polling.enable. poll_switch() goes through interface list and enabled/disables polling. A message that kern.polling.enable is deprecated is printed. Detailed driver changes: - On attach driver announces IFCAP_POLLING in if_capabilities, but not in if_capenable. - On detach driver calls ether_poll_deregister() if polling is enabled. - In polling handler driver obtains its lock and checks IFF_DRV_RUNNING flag. If there is no, then unlocks and returns. - In ioctl handler driver checks for IFCAP_POLLING flag requested to be set or cleared. Driver first calls ether_poll_[de]register(), then obtains driver lock and [dis/en]ables interrupts. - In interrupt handler driver checks IFCAP_POLLING flag in if_capenable. If present, then returns.This is important to protect from spurious interrupts. Reviewed by: ru, sam, jhb
2005-10-01 18:56:19 +00:00
ifp->if_capenable = ifp->if_capabilities;
#ifdef DEVICE_POLLING
ifp->if_capabilities |= IFCAP_POLLING;
#endif
/* Hook interrupt last to avoid having to lock softc */
error = bus_setup_intr(dev, sc->sis_res[1], INTR_TYPE_NET | INTR_MPSAFE,
NULL, sis_intr, sc, &sc->sis_intrhand);
Clean up locking and resource management for pci/if_* - Remove locking of the softc in the attach method, instead depending on bus_setup_intr being at the end of attach (delaying interrupt enable until after ether_ifattach is called) - Call *_detach directly in the error case of attach, depending on checking in detach to only free resources that were allocated. This puts all resource freeing in one place, avoiding thinkos that lead to memory leaks. - Add bus_child_present check to calls to *_stop in the detach method to be sure hw is present before touching its registers. - Remove bzero softc calls since device_t should do this for us. - dc: move interrupt allocation back where it was before. It was unnecessary to move it. This reverts part of 1.88 - rl: move irq allocation before ether_ifattach. Problems might have been caused by allocating the irq after enabling interrupts on the card. - rl: call rl_stop before ether_ifdetach - sf: call sf_stop before ether_ifdetach - sis: add missed free of sis_tag - sis: check errors from tag creation - sis: move dmamem_alloc and dmamap_load to happen at same time as tag creation - sk: remove duplicate initialization of sk_dev - ste: add missed bus_generic_detach - ti: call ti_stop before ether_ifdetach - ti: add missed error setting in ti_rdata alloc failure - vr: add missed error setting in I/O, memory mapping cases - xl: add missed error setting in I/O, memory mapping cases - xl: remove multi-level goto on attach failure - xl: move dmamem_alloc and dmamap_load to happen at same time as tag creation - Calls to free(9) are unconditional because it is valid to call free with a null pointer. Reviewed by: imp, mdodd
2003-03-31 17:29:43 +00:00
if (error) {
device_printf(dev, "couldn't set up irq\n");
ether_ifdetach(ifp);
Clean up locking and resource management for pci/if_* - Remove locking of the softc in the attach method, instead depending on bus_setup_intr being at the end of attach (delaying interrupt enable until after ether_ifattach is called) - Call *_detach directly in the error case of attach, depending on checking in detach to only free resources that were allocated. This puts all resource freeing in one place, avoiding thinkos that lead to memory leaks. - Add bus_child_present check to calls to *_stop in the detach method to be sure hw is present before touching its registers. - Remove bzero softc calls since device_t should do this for us. - dc: move interrupt allocation back where it was before. It was unnecessary to move it. This reverts part of 1.88 - rl: move irq allocation before ether_ifattach. Problems might have been caused by allocating the irq after enabling interrupts on the card. - rl: call rl_stop before ether_ifdetach - sf: call sf_stop before ether_ifdetach - sis: add missed free of sis_tag - sis: check errors from tag creation - sis: move dmamem_alloc and dmamap_load to happen at same time as tag creation - sk: remove duplicate initialization of sk_dev - ste: add missed bus_generic_detach - ti: call ti_stop before ether_ifdetach - ti: add missed error setting in ti_rdata alloc failure - vr: add missed error setting in I/O, memory mapping cases - xl: add missed error setting in I/O, memory mapping cases - xl: remove multi-level goto on attach failure - xl: move dmamem_alloc and dmamap_load to happen at same time as tag creation - Calls to free(9) are unconditional because it is valid to call free with a null pointer. Reviewed by: imp, mdodd
2003-03-31 17:29:43 +00:00
goto fail;
}
fail:
Clean up locking and resource management for pci/if_* - Remove locking of the softc in the attach method, instead depending on bus_setup_intr being at the end of attach (delaying interrupt enable until after ether_ifattach is called) - Call *_detach directly in the error case of attach, depending on checking in detach to only free resources that were allocated. This puts all resource freeing in one place, avoiding thinkos that lead to memory leaks. - Add bus_child_present check to calls to *_stop in the detach method to be sure hw is present before touching its registers. - Remove bzero softc calls since device_t should do this for us. - dc: move interrupt allocation back where it was before. It was unnecessary to move it. This reverts part of 1.88 - rl: move irq allocation before ether_ifattach. Problems might have been caused by allocating the irq after enabling interrupts on the card. - rl: call rl_stop before ether_ifdetach - sf: call sf_stop before ether_ifdetach - sis: add missed free of sis_tag - sis: check errors from tag creation - sis: move dmamem_alloc and dmamap_load to happen at same time as tag creation - sk: remove duplicate initialization of sk_dev - ste: add missed bus_generic_detach - ti: call ti_stop before ether_ifdetach - ti: add missed error setting in ti_rdata alloc failure - vr: add missed error setting in I/O, memory mapping cases - xl: add missed error setting in I/O, memory mapping cases - xl: remove multi-level goto on attach failure - xl: move dmamem_alloc and dmamap_load to happen at same time as tag creation - Calls to free(9) are unconditional because it is valid to call free with a null pointer. Reviewed by: imp, mdodd
2003-03-31 17:29:43 +00:00
if (error)
sis_detach(dev);
2010-09-01 18:39:35 +00:00
return (error);
}
/*
* Shutdown hardware and free up resources. This can be called any
* time after the mutex has been initialized. It is called in both
* the error case in attach and the normal detach case so it needs
* to be careful about only freeing resources that have actually been
* allocated.
*/
static int
sis_detach(device_t dev)
{
struct sis_softc *sc;
struct ifnet *ifp;
sc = device_get_softc(dev);
KASSERT(mtx_initialized(&sc->sis_mtx), ("sis mutex not initialized"));
ifp = sc->sis_ifp;
Big polling(4) cleanup. o Axe poll in trap. o Axe IFF_POLLING flag from if_flags. o Rework revision 1.21 (Giant removal), in such a way that poll_mtx is not dropped during call to polling handler. This fixes problem with idle polling. o Make registration and deregistration from polling in a functional way, insted of next tick/interrupt. o Obsolete kern.polling.enable. Polling is turned on/off with ifconfig. Detailed kern_poll.c changes: - Remove polling handler flags, introduced in 1.21. The are not needed now. - Forget and do not check if_flags, if_capenable and if_drv_flags. - Call all registered polling handlers unconditionally. - Do not drop poll_mtx, when entering polling handlers. - In ether_poll() NET_LOCK_GIANT prior to locking poll_mtx. - In netisr_poll() axe the block, where polling code asks drivers to unregister. - In netisr_poll() and ether_poll() do polling always, if any handlers are present. - In ether_poll_[de]register() remove a lot of error hiding code. Assert that arguments are correct, instead. - In ether_poll_[de]register() use standard return values in case of error or success. - Introduce poll_switch() that is a sysctl handler for kern.polling.enable. poll_switch() goes through interface list and enabled/disables polling. A message that kern.polling.enable is deprecated is printed. Detailed driver changes: - On attach driver announces IFCAP_POLLING in if_capabilities, but not in if_capenable. - On detach driver calls ether_poll_deregister() if polling is enabled. - In polling handler driver obtains its lock and checks IFF_DRV_RUNNING flag. If there is no, then unlocks and returns. - In ioctl handler driver checks for IFCAP_POLLING flag requested to be set or cleared. Driver first calls ether_poll_[de]register(), then obtains driver lock and [dis/en]ables interrupts. - In interrupt handler driver checks IFCAP_POLLING flag in if_capenable. If present, then returns.This is important to protect from spurious interrupts. Reviewed by: ru, sam, jhb
2005-10-01 18:56:19 +00:00
#ifdef DEVICE_POLLING
if (ifp->if_capenable & IFCAP_POLLING)
ether_poll_deregister(ifp);
#endif
2003-07-27 14:38:54 +00:00
/* These should only be active if attach succeeded. */
if (device_is_attached(dev)) {
SIS_LOCK(sc);
sis_stop(sc);
SIS_UNLOCK(sc);
callout_drain(&sc->sis_stat_ch);
Clean up locking and resource management for pci/if_* - Remove locking of the softc in the attach method, instead depending on bus_setup_intr being at the end of attach (delaying interrupt enable until after ether_ifattach is called) - Call *_detach directly in the error case of attach, depending on checking in detach to only free resources that were allocated. This puts all resource freeing in one place, avoiding thinkos that lead to memory leaks. - Add bus_child_present check to calls to *_stop in the detach method to be sure hw is present before touching its registers. - Remove bzero softc calls since device_t should do this for us. - dc: move interrupt allocation back where it was before. It was unnecessary to move it. This reverts part of 1.88 - rl: move irq allocation before ether_ifattach. Problems might have been caused by allocating the irq after enabling interrupts on the card. - rl: call rl_stop before ether_ifdetach - sf: call sf_stop before ether_ifdetach - sis: add missed free of sis_tag - sis: check errors from tag creation - sis: move dmamem_alloc and dmamap_load to happen at same time as tag creation - sk: remove duplicate initialization of sk_dev - ste: add missed bus_generic_detach - ti: call ti_stop before ether_ifdetach - ti: add missed error setting in ti_rdata alloc failure - vr: add missed error setting in I/O, memory mapping cases - xl: add missed error setting in I/O, memory mapping cases - xl: remove multi-level goto on attach failure - xl: move dmamem_alloc and dmamap_load to happen at same time as tag creation - Calls to free(9) are unconditional because it is valid to call free with a null pointer. Reviewed by: imp, mdodd
2003-03-31 17:29:43 +00:00
ether_ifdetach(ifp);
}
if (sc->sis_miibus)
device_delete_child(dev, sc->sis_miibus);
bus_generic_detach(dev);
Clean up locking and resource management for pci/if_* - Remove locking of the softc in the attach method, instead depending on bus_setup_intr being at the end of attach (delaying interrupt enable until after ether_ifattach is called) - Call *_detach directly in the error case of attach, depending on checking in detach to only free resources that were allocated. This puts all resource freeing in one place, avoiding thinkos that lead to memory leaks. - Add bus_child_present check to calls to *_stop in the detach method to be sure hw is present before touching its registers. - Remove bzero softc calls since device_t should do this for us. - dc: move interrupt allocation back where it was before. It was unnecessary to move it. This reverts part of 1.88 - rl: move irq allocation before ether_ifattach. Problems might have been caused by allocating the irq after enabling interrupts on the card. - rl: call rl_stop before ether_ifdetach - sf: call sf_stop before ether_ifdetach - sis: add missed free of sis_tag - sis: check errors from tag creation - sis: move dmamem_alloc and dmamap_load to happen at same time as tag creation - sk: remove duplicate initialization of sk_dev - ste: add missed bus_generic_detach - ti: call ti_stop before ether_ifdetach - ti: add missed error setting in ti_rdata alloc failure - vr: add missed error setting in I/O, memory mapping cases - xl: add missed error setting in I/O, memory mapping cases - xl: remove multi-level goto on attach failure - xl: move dmamem_alloc and dmamap_load to happen at same time as tag creation - Calls to free(9) are unconditional because it is valid to call free with a null pointer. Reviewed by: imp, mdodd
2003-03-31 17:29:43 +00:00
if (sc->sis_intrhand)
bus_teardown_intr(dev, sc->sis_res[1], sc->sis_intrhand);
bus_release_resources(dev, sis_res_spec, sc->sis_res);
Clean up locking and resource management for pci/if_* - Remove locking of the softc in the attach method, instead depending on bus_setup_intr being at the end of attach (delaying interrupt enable until after ether_ifattach is called) - Call *_detach directly in the error case of attach, depending on checking in detach to only free resources that were allocated. This puts all resource freeing in one place, avoiding thinkos that lead to memory leaks. - Add bus_child_present check to calls to *_stop in the detach method to be sure hw is present before touching its registers. - Remove bzero softc calls since device_t should do this for us. - dc: move interrupt allocation back where it was before. It was unnecessary to move it. This reverts part of 1.88 - rl: move irq allocation before ether_ifattach. Problems might have been caused by allocating the irq after enabling interrupts on the card. - rl: call rl_stop before ether_ifdetach - sf: call sf_stop before ether_ifdetach - sis: add missed free of sis_tag - sis: check errors from tag creation - sis: move dmamem_alloc and dmamap_load to happen at same time as tag creation - sk: remove duplicate initialization of sk_dev - ste: add missed bus_generic_detach - ti: call ti_stop before ether_ifdetach - ti: add missed error setting in ti_rdata alloc failure - vr: add missed error setting in I/O, memory mapping cases - xl: add missed error setting in I/O, memory mapping cases - xl: remove multi-level goto on attach failure - xl: move dmamem_alloc and dmamap_load to happen at same time as tag creation - Calls to free(9) are unconditional because it is valid to call free with a null pointer. Reviewed by: imp, mdodd
2003-03-31 17:29:43 +00:00
if (ifp)
if_free(ifp);
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
sis_dma_free(sc);
mtx_destroy(&sc->sis_mtx);
return (0);
}
struct sis_dmamap_arg {
bus_addr_t sis_busaddr;
};
static void
sis_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
{
struct sis_dmamap_arg *ctx;
if (error != 0)
return;
KASSERT(nsegs == 1, ("%s: %d segments returned!", __func__, nsegs));
ctx = (struct sis_dmamap_arg *)arg;
ctx->sis_busaddr = segs[0].ds_addr;
}
static int
sis_dma_ring_alloc(struct sis_softc *sc, bus_size_t alignment,
bus_size_t maxsize, bus_dma_tag_t *tag, uint8_t **ring, bus_dmamap_t *map,
bus_addr_t *paddr, const char *msg)
{
struct sis_dmamap_arg ctx;
int error;
error = bus_dma_tag_create(sc->sis_parent_tag, alignment, 0,
BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, maxsize, 1,
maxsize, 0, NULL, NULL, tag);
if (error != 0) {
device_printf(sc->sis_dev,
"could not create %s dma tag\n", msg);
return (ENOMEM);
Clean up locking and resource management for pci/if_* - Remove locking of the softc in the attach method, instead depending on bus_setup_intr being at the end of attach (delaying interrupt enable until after ether_ifattach is called) - Call *_detach directly in the error case of attach, depending on checking in detach to only free resources that were allocated. This puts all resource freeing in one place, avoiding thinkos that lead to memory leaks. - Add bus_child_present check to calls to *_stop in the detach method to be sure hw is present before touching its registers. - Remove bzero softc calls since device_t should do this for us. - dc: move interrupt allocation back where it was before. It was unnecessary to move it. This reverts part of 1.88 - rl: move irq allocation before ether_ifattach. Problems might have been caused by allocating the irq after enabling interrupts on the card. - rl: call rl_stop before ether_ifdetach - sf: call sf_stop before ether_ifdetach - sis: add missed free of sis_tag - sis: check errors from tag creation - sis: move dmamem_alloc and dmamap_load to happen at same time as tag creation - sk: remove duplicate initialization of sk_dev - ste: add missed bus_generic_detach - ti: call ti_stop before ether_ifdetach - ti: add missed error setting in ti_rdata alloc failure - vr: add missed error setting in I/O, memory mapping cases - xl: add missed error setting in I/O, memory mapping cases - xl: remove multi-level goto on attach failure - xl: move dmamem_alloc and dmamap_load to happen at same time as tag creation - Calls to free(9) are unconditional because it is valid to call free with a null pointer. Reviewed by: imp, mdodd
2003-03-31 17:29:43 +00:00
}
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
/* Allocate DMA'able memory for ring. */
error = bus_dmamem_alloc(*tag, (void **)ring,
BUS_DMA_NOWAIT | BUS_DMA_ZERO | BUS_DMA_COHERENT, map);
if (error != 0) {
device_printf(sc->sis_dev,
"could not allocate DMA'able memory for %s\n", msg);
return (ENOMEM);
}
/* Load the address of the ring. */
ctx.sis_busaddr = 0;
error = bus_dmamap_load(*tag, *map, *ring, maxsize, sis_dmamap_cb,
&ctx, BUS_DMA_NOWAIT);
if (error != 0) {
device_printf(sc->sis_dev,
"could not load DMA'able memory for %s\n", msg);
return (ENOMEM);
Clean up locking and resource management for pci/if_* - Remove locking of the softc in the attach method, instead depending on bus_setup_intr being at the end of attach (delaying interrupt enable until after ether_ifattach is called) - Call *_detach directly in the error case of attach, depending on checking in detach to only free resources that were allocated. This puts all resource freeing in one place, avoiding thinkos that lead to memory leaks. - Add bus_child_present check to calls to *_stop in the detach method to be sure hw is present before touching its registers. - Remove bzero softc calls since device_t should do this for us. - dc: move interrupt allocation back where it was before. It was unnecessary to move it. This reverts part of 1.88 - rl: move irq allocation before ether_ifattach. Problems might have been caused by allocating the irq after enabling interrupts on the card. - rl: call rl_stop before ether_ifdetach - sf: call sf_stop before ether_ifdetach - sis: add missed free of sis_tag - sis: check errors from tag creation - sis: move dmamem_alloc and dmamap_load to happen at same time as tag creation - sk: remove duplicate initialization of sk_dev - ste: add missed bus_generic_detach - ti: call ti_stop before ether_ifdetach - ti: add missed error setting in ti_rdata alloc failure - vr: add missed error setting in I/O, memory mapping cases - xl: add missed error setting in I/O, memory mapping cases - xl: remove multi-level goto on attach failure - xl: move dmamem_alloc and dmamap_load to happen at same time as tag creation - Calls to free(9) are unconditional because it is valid to call free with a null pointer. Reviewed by: imp, mdodd
2003-03-31 17:29:43 +00:00
}
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
*paddr = ctx.sis_busaddr;
return (0);
}
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
static int
sis_dma_alloc(struct sis_softc *sc)
{
struct sis_rxdesc *rxd;
struct sis_txdesc *txd;
int error, i;
/* Allocate the parent bus DMA tag appropriate for PCI. */
error = bus_dma_tag_create(bus_get_dma_tag(sc->sis_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->sis_parent_tag);
if (error != 0) {
device_printf(sc->sis_dev,
"could not allocate parent dma tag\n");
return (ENOMEM);
}
/* Create RX ring. */
error = sis_dma_ring_alloc(sc, SIS_DESC_ALIGN, SIS_RX_LIST_SZ,
&sc->sis_rx_list_tag, (uint8_t **)&sc->sis_rx_list,
&sc->sis_rx_list_map, &sc->sis_rx_paddr, "RX ring");
if (error)
return (error);
/* Create TX ring. */
error = sis_dma_ring_alloc(sc, SIS_DESC_ALIGN, SIS_TX_LIST_SZ,
&sc->sis_tx_list_tag, (uint8_t **)&sc->sis_tx_list,
&sc->sis_tx_list_map, &sc->sis_tx_paddr, "TX ring");
if (error)
return (error);
/* Create tag for RX mbufs. */
error = bus_dma_tag_create(sc->sis_parent_tag, SIS_RX_BUF_ALIGN, 0,
BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, MCLBYTES, 1,
MCLBYTES, 0, NULL, NULL, &sc->sis_rx_tag);
if (error) {
device_printf(sc->sis_dev, "could not allocate RX dma tag\n");
return (error);
}
/* Create tag for TX mbufs. */
error = bus_dma_tag_create(sc->sis_parent_tag, 1, 0,
BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
MCLBYTES * SIS_MAXTXSEGS, SIS_MAXTXSEGS, MCLBYTES, 0, NULL, NULL,
&sc->sis_tx_tag);
if (error) {
device_printf(sc->sis_dev, "could not allocate TX dma tag\n");
return (error);
}
/* Create DMA maps for RX buffers. */
error = bus_dmamap_create(sc->sis_rx_tag, 0, &sc->sis_rx_sparemap);
if (error) {
device_printf(sc->sis_dev,
"can't create spare DMA map for RX\n");
return (error);
}
for (i = 0; i < SIS_RX_LIST_CNT; i++) {
rxd = &sc->sis_rxdesc[i];
rxd->rx_m = NULL;
error = bus_dmamap_create(sc->sis_rx_tag, 0, &rxd->rx_dmamap);
if (error) {
device_printf(sc->sis_dev,
"can't create DMA map for RX\n");
return (error);
}
}
/* Create DMA maps for TX buffers. */
for (i = 0; i < SIS_TX_LIST_CNT; i++) {
txd = &sc->sis_txdesc[i];
txd->tx_m = NULL;
error = bus_dmamap_create(sc->sis_tx_tag, 0, &txd->tx_dmamap);
if (error) {
device_printf(sc->sis_dev,
"can't create DMA map for TX\n");
return (error);
}
}
2010-09-01 18:39:35 +00:00
return (0);
}
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
static void
sis_dma_free(struct sis_softc *sc)
{
struct sis_rxdesc *rxd;
struct sis_txdesc *txd;
int i;
/* Destroy DMA maps for RX buffers. */
for (i = 0; i < SIS_RX_LIST_CNT; i++) {
rxd = &sc->sis_rxdesc[i];
if (rxd->rx_dmamap)
bus_dmamap_destroy(sc->sis_rx_tag, rxd->rx_dmamap);
}
if (sc->sis_rx_sparemap)
bus_dmamap_destroy(sc->sis_rx_tag, sc->sis_rx_sparemap);
/* Destroy DMA maps for TX buffers. */
for (i = 0; i < SIS_TX_LIST_CNT; i++) {
txd = &sc->sis_txdesc[i];
if (txd->tx_dmamap)
bus_dmamap_destroy(sc->sis_tx_tag, txd->tx_dmamap);
}
if (sc->sis_rx_tag)
bus_dma_tag_destroy(sc->sis_rx_tag);
if (sc->sis_tx_tag)
bus_dma_tag_destroy(sc->sis_tx_tag);
/* Destroy RX ring. */
if (sc->sis_rx_list_map)
bus_dmamap_unload(sc->sis_rx_list_tag, sc->sis_rx_list_map);
if (sc->sis_rx_list_map && sc->sis_rx_list)
bus_dmamem_free(sc->sis_rx_list_tag, sc->sis_rx_list,
sc->sis_rx_list_map);
if (sc->sis_rx_list_tag)
bus_dma_tag_destroy(sc->sis_rx_list_tag);
/* Destroy TX ring. */
if (sc->sis_tx_list_map)
bus_dmamap_unload(sc->sis_tx_list_tag, sc->sis_tx_list_map);
if (sc->sis_tx_list_map && sc->sis_tx_list)
bus_dmamem_free(sc->sis_tx_list_tag, sc->sis_tx_list,
sc->sis_tx_list_map);
if (sc->sis_tx_list_tag)
bus_dma_tag_destroy(sc->sis_tx_list_tag);
/* Destroy the parent tag. */
if (sc->sis_parent_tag)
bus_dma_tag_destroy(sc->sis_parent_tag);
}
/*
* Initialize the TX and RX descriptors and allocate mbufs for them. Note that
* we arrange the descriptors in a closed ring, so that the last descriptor
* points back to the first.
*/
static int
sis_ring_init(struct sis_softc *sc)
{
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
struct sis_rxdesc *rxd;
struct sis_txdesc *txd;
bus_addr_t next;
int error, i;
bzero(&sc->sis_tx_list[0], SIS_TX_LIST_SZ);
for (i = 0; i < SIS_TX_LIST_CNT; i++) {
txd = &sc->sis_txdesc[i];
txd->tx_m = NULL;
if (i == SIS_TX_LIST_CNT - 1)
next = SIS_TX_RING_ADDR(sc, 0);
else
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
next = SIS_TX_RING_ADDR(sc, i + 1);
sc->sis_tx_list[i].sis_next = htole32(SIS_ADDR_LO(next));
}
sc->sis_tx_prod = sc->sis_tx_cons = sc->sis_tx_cnt = 0;
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
bus_dmamap_sync(sc->sis_tx_list_tag, sc->sis_tx_list_map,
BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
sc->sis_rx_cons = 0;
bzero(&sc->sis_rx_list[0], SIS_RX_LIST_SZ);
for (i = 0; i < SIS_RX_LIST_CNT; i++) {
rxd = &sc->sis_rxdesc[i];
rxd->rx_desc = &sc->sis_rx_list[i];
if (i == SIS_RX_LIST_CNT - 1)
next = SIS_RX_RING_ADDR(sc, 0);
else
next = SIS_RX_RING_ADDR(sc, i + 1);
rxd->rx_desc->sis_next = htole32(SIS_ADDR_LO(next));
error = sis_newbuf(sc, rxd);
if (error)
2010-09-01 18:39:35 +00:00
return (error);
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
}
bus_dmamap_sync(sc->sis_rx_list_tag, sc->sis_rx_list_map,
BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2010-09-01 18:39:35 +00:00
return (0);
}
/*
* Initialize an RX descriptor and attach an MBUF cluster.
*/
static int
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
sis_newbuf(struct sis_softc *sc, struct sis_rxdesc *rxd)
{
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
struct mbuf *m;
bus_dma_segment_t segs[1];
bus_dmamap_t map;
int nsegs;
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
m = m_getcl(M_DONTWAIT, MT_DATA, M_PKTHDR);
if (m == NULL)
return (ENOBUFS);
m->m_len = m->m_pkthdr.len = SIS_RXLEN;
#ifndef __NO_STRICT_ALIGNMENT
m_adj(m, SIS_RX_BUF_ALIGN);
#endif
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
if (bus_dmamap_load_mbuf_sg(sc->sis_rx_tag, sc->sis_rx_sparemap, m,
segs, &nsegs, 0) != 0) {
m_freem(m);
return (ENOBUFS);
}
KASSERT(nsegs == 1, ("%s: %d segments returned!", __func__, nsegs));
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
if (rxd->rx_m != NULL) {
bus_dmamap_sync(sc->sis_rx_tag, rxd->rx_dmamap,
BUS_DMASYNC_POSTREAD);
bus_dmamap_unload(sc->sis_rx_tag, rxd->rx_dmamap);
}
map = rxd->rx_dmamap;
rxd->rx_dmamap = sc->sis_rx_sparemap;
sc->sis_rx_sparemap = map;
bus_dmamap_sync(sc->sis_rx_tag, rxd->rx_dmamap, BUS_DMASYNC_PREREAD);
rxd->rx_m = m;
rxd->rx_desc->sis_cmdsts = htole32(SIS_RXLEN);
rxd->rx_desc->sis_ptr = htole32(SIS_ADDR_LO(segs[0].ds_addr));
return (0);
}
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
static __inline void
sis_discard_rxbuf(struct sis_rxdesc *rxd)
{
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
rxd->rx_desc->sis_cmdsts = htole32(SIS_RXLEN);
}
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
#ifndef __NO_STRICT_ALIGNMENT
static __inline void
sis_fixup_rx(struct mbuf *m)
{
uint16_t *src, *dst;
int i;
src = mtod(m, uint16_t *);
dst = src - (SIS_RX_BUF_ALIGN - ETHER_ALIGN) / sizeof(*src);
for (i = 0; i < (m->m_len / sizeof(uint16_t) + 1); i++)
*dst++ = *src++;
m->m_data -= SIS_RX_BUF_ALIGN - ETHER_ALIGN;
}
#endif
/*
* A frame has been uploaded: pass the resulting mbuf chain up to
* the higher level protocols.
*/
static int
sis_rxeof(struct sis_softc *sc)
{
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
struct mbuf *m;
struct ifnet *ifp;
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
struct sis_rxdesc *rxd;
struct sis_desc *cur_rx;
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
int prog, rx_cons, rx_npkts = 0, total_len;
uint32_t rxstat;
SIS_LOCK_ASSERT(sc);
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
bus_dmamap_sync(sc->sis_rx_list_tag, sc->sis_rx_list_map,
BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
rx_cons = sc->sis_rx_cons;
ifp = sc->sis_ifp;
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
for (prog = 0; (ifp->if_drv_flags & IFF_DRV_RUNNING) != 0;
SIS_INC(rx_cons, SIS_RX_LIST_CNT), prog++) {
Device Polling code for -current. Non-SMP, i386-only, no polling in the idle loop at the moment. To use this code you must compile a kernel with options DEVICE_POLLING and at runtime enable polling with sysctl kern.polling.enable=1 The percentage of CPU reserved to userland can be set with sysctl kern.polling.user_frac=NN (default is 50) while the remainder is used by polling device drivers and netisr's. These are the only two variables that you should need to touch. There are a few more parameters in kern.polling but the default values are adequate for all purposes. See the code in kern_poll.c for more details on them. Polling in the idle loop will be implemented shortly by introducing a kernel thread which does the job. Until then, the amount of CPU dedicated to polling will never exceed (100-user_frac). The equivalent (actually, better) code for -stable is at http://info.iet.unipi.it/~luigi/polling/ and also supports polling in the idle loop. NOTE to Alpha developers: There is really nothing in this code that is i386-specific. If you move the 2 lines supporting the new option from sys/conf/{files,options}.i386 to sys/conf/{files,options} I am pretty sure that this should work on the Alpha as well, just that I do not have a suitable test box to try it. If someone feels like trying it, I would appreciate it. NOTE to other developers: sure some things could be done better, and as always I am open to constructive criticism, which a few of you have already given and I greatly appreciated. However, before proposing radical architectural changes, please take some time to possibly try out this code, or at the very least read the comments in kern_poll.c, especially re. the reason why I am using a soft netisr and cannot (I believe) replace it with a simple timeout. Quick description of files touched by this commit: sys/conf/files.i386 new file kern/kern_poll.c sys/conf/options.i386 new option sys/i386/i386/trap.c poll in trap (disabled by default) sys/kern/kern_clock.c initialization and hardclock hooks. sys/kern/kern_intr.c minor swi_net changes sys/kern/kern_poll.c the bulk of the code. sys/net/if.h new flag sys/net/if_var.h declaration for functions used in device drivers. sys/net/netisr.h NETISR_POLL sys/dev/fxp/if_fxp.c sys/dev/fxp/if_fxpvar.h sys/pci/if_dc.c sys/pci/if_dcreg.h sys/pci/if_sis.c sys/pci/if_sisreg.h device driver modifications
2001-12-14 17:56:12 +00:00
#ifdef DEVICE_POLLING
Big polling(4) cleanup. o Axe poll in trap. o Axe IFF_POLLING flag from if_flags. o Rework revision 1.21 (Giant removal), in such a way that poll_mtx is not dropped during call to polling handler. This fixes problem with idle polling. o Make registration and deregistration from polling in a functional way, insted of next tick/interrupt. o Obsolete kern.polling.enable. Polling is turned on/off with ifconfig. Detailed kern_poll.c changes: - Remove polling handler flags, introduced in 1.21. The are not needed now. - Forget and do not check if_flags, if_capenable and if_drv_flags. - Call all registered polling handlers unconditionally. - Do not drop poll_mtx, when entering polling handlers. - In ether_poll() NET_LOCK_GIANT prior to locking poll_mtx. - In netisr_poll() axe the block, where polling code asks drivers to unregister. - In netisr_poll() and ether_poll() do polling always, if any handlers are present. - In ether_poll_[de]register() remove a lot of error hiding code. Assert that arguments are correct, instead. - In ether_poll_[de]register() use standard return values in case of error or success. - Introduce poll_switch() that is a sysctl handler for kern.polling.enable. poll_switch() goes through interface list and enabled/disables polling. A message that kern.polling.enable is deprecated is printed. Detailed driver changes: - On attach driver announces IFCAP_POLLING in if_capabilities, but not in if_capenable. - On detach driver calls ether_poll_deregister() if polling is enabled. - In polling handler driver obtains its lock and checks IFF_DRV_RUNNING flag. If there is no, then unlocks and returns. - In ioctl handler driver checks for IFCAP_POLLING flag requested to be set or cleared. Driver first calls ether_poll_[de]register(), then obtains driver lock and [dis/en]ables interrupts. - In interrupt handler driver checks IFCAP_POLLING flag in if_capenable. If present, then returns.This is important to protect from spurious interrupts. Reviewed by: ru, sam, jhb
2005-10-01 18:56:19 +00:00
if (ifp->if_capenable & IFCAP_POLLING) {
Device Polling code for -current. Non-SMP, i386-only, no polling in the idle loop at the moment. To use this code you must compile a kernel with options DEVICE_POLLING and at runtime enable polling with sysctl kern.polling.enable=1 The percentage of CPU reserved to userland can be set with sysctl kern.polling.user_frac=NN (default is 50) while the remainder is used by polling device drivers and netisr's. These are the only two variables that you should need to touch. There are a few more parameters in kern.polling but the default values are adequate for all purposes. See the code in kern_poll.c for more details on them. Polling in the idle loop will be implemented shortly by introducing a kernel thread which does the job. Until then, the amount of CPU dedicated to polling will never exceed (100-user_frac). The equivalent (actually, better) code for -stable is at http://info.iet.unipi.it/~luigi/polling/ and also supports polling in the idle loop. NOTE to Alpha developers: There is really nothing in this code that is i386-specific. If you move the 2 lines supporting the new option from sys/conf/{files,options}.i386 to sys/conf/{files,options} I am pretty sure that this should work on the Alpha as well, just that I do not have a suitable test box to try it. If someone feels like trying it, I would appreciate it. NOTE to other developers: sure some things could be done better, and as always I am open to constructive criticism, which a few of you have already given and I greatly appreciated. However, before proposing radical architectural changes, please take some time to possibly try out this code, or at the very least read the comments in kern_poll.c, especially re. the reason why I am using a soft netisr and cannot (I believe) replace it with a simple timeout. Quick description of files touched by this commit: sys/conf/files.i386 new file kern/kern_poll.c sys/conf/options.i386 new option sys/i386/i386/trap.c poll in trap (disabled by default) sys/kern/kern_clock.c initialization and hardclock hooks. sys/kern/kern_intr.c minor swi_net changes sys/kern/kern_poll.c the bulk of the code. sys/net/if.h new flag sys/net/if_var.h declaration for functions used in device drivers. sys/net/netisr.h NETISR_POLL sys/dev/fxp/if_fxp.c sys/dev/fxp/if_fxpvar.h sys/pci/if_dc.c sys/pci/if_dcreg.h sys/pci/if_sis.c sys/pci/if_sisreg.h device driver modifications
2001-12-14 17:56:12 +00:00
if (sc->rxcycles <= 0)
break;
sc->rxcycles--;
}
Big polling(4) cleanup. o Axe poll in trap. o Axe IFF_POLLING flag from if_flags. o Rework revision 1.21 (Giant removal), in such a way that poll_mtx is not dropped during call to polling handler. This fixes problem with idle polling. o Make registration and deregistration from polling in a functional way, insted of next tick/interrupt. o Obsolete kern.polling.enable. Polling is turned on/off with ifconfig. Detailed kern_poll.c changes: - Remove polling handler flags, introduced in 1.21. The are not needed now. - Forget and do not check if_flags, if_capenable and if_drv_flags. - Call all registered polling handlers unconditionally. - Do not drop poll_mtx, when entering polling handlers. - In ether_poll() NET_LOCK_GIANT prior to locking poll_mtx. - In netisr_poll() axe the block, where polling code asks drivers to unregister. - In netisr_poll() and ether_poll() do polling always, if any handlers are present. - In ether_poll_[de]register() remove a lot of error hiding code. Assert that arguments are correct, instead. - In ether_poll_[de]register() use standard return values in case of error or success. - Introduce poll_switch() that is a sysctl handler for kern.polling.enable. poll_switch() goes through interface list and enabled/disables polling. A message that kern.polling.enable is deprecated is printed. Detailed driver changes: - On attach driver announces IFCAP_POLLING in if_capabilities, but not in if_capenable. - On detach driver calls ether_poll_deregister() if polling is enabled. - In polling handler driver obtains its lock and checks IFF_DRV_RUNNING flag. If there is no, then unlocks and returns. - In ioctl handler driver checks for IFCAP_POLLING flag requested to be set or cleared. Driver first calls ether_poll_[de]register(), then obtains driver lock and [dis/en]ables interrupts. - In interrupt handler driver checks IFCAP_POLLING flag in if_capenable. If present, then returns.This is important to protect from spurious interrupts. Reviewed by: ru, sam, jhb
2005-10-01 18:56:19 +00:00
#endif
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
cur_rx = &sc->sis_rx_list[rx_cons];
rxstat = le32toh(cur_rx->sis_cmdsts);
if ((rxstat & SIS_CMDSTS_OWN) == 0)
break;
rxd = &sc->sis_rxdesc[rx_cons];
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
total_len = (rxstat & SIS_CMDSTS_BUFLEN) - ETHER_CRC_LEN;
if ((ifp->if_capenable & IFCAP_VLAN_MTU) != 0 &&
total_len <= (ETHER_MAX_LEN + ETHER_VLAN_ENCAP_LEN -
ETHER_CRC_LEN))
rxstat &= ~SIS_RXSTAT_GIANT;
if (SIS_RXSTAT_ERROR(rxstat) != 0) {
ifp->if_ierrors++;
if (rxstat & SIS_RXSTAT_COLL)
ifp->if_collisions++;
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
sis_discard_rxbuf(rxd);
continue;
}
/* Add a new receive buffer to the ring. */
m = rxd->rx_m;
if (sis_newbuf(sc, rxd) != 0) {
ifp->if_iqdrops++;
sis_discard_rxbuf(rxd);
continue;
}
/* No errors; receive the packet. */
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
m->m_pkthdr.len = m->m_len = total_len;
#ifndef __NO_STRICT_ALIGNMENT
/*
* On architectures without alignment problems we try to
* allocate a new buffer for the receive ring, and pass up
* the one where the packet is already, saving the expensive
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
* copy operation.
*/
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
sis_fixup_rx(m);
#endif
ifp->if_ipackets++;
m->m_pkthdr.rcvif = ifp;
SIS_UNLOCK(sc);
(*ifp->if_input)(ifp, m);
SIS_LOCK(sc);
rx_npkts++;
}
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
if (prog > 0) {
sc->sis_rx_cons = rx_cons;
bus_dmamap_sync(sc->sis_rx_list_tag, sc->sis_rx_list_map,
BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
}
return (rx_npkts);
}
/*
* A frame was downloaded to the chip. It's safe for us to clean up
* the list buffers.
*/
static void
sis_txeof(struct sis_softc *sc)
{
struct ifnet *ifp;
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
struct sis_desc *cur_tx;
struct sis_txdesc *txd;
uint32_t cons, txstat;
SIS_LOCK_ASSERT(sc);
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
cons = sc->sis_tx_cons;
if (cons == sc->sis_tx_prod)
return;
ifp = sc->sis_ifp;
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
bus_dmamap_sync(sc->sis_tx_list_tag, sc->sis_tx_list_map,
BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
/*
* Go through our tx list and free mbufs for those
* frames that have been transmitted.
*/
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
for (; cons != sc->sis_tx_prod; SIS_INC(cons, SIS_TX_LIST_CNT)) {
cur_tx = &sc->sis_tx_list[cons];
txstat = le32toh(cur_tx->sis_cmdsts);
if ((txstat & SIS_CMDSTS_OWN) != 0)
break;
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
txd = &sc->sis_txdesc[cons];
if (txd->tx_m != NULL) {
bus_dmamap_sync(sc->sis_tx_tag, txd->tx_dmamap,
BUS_DMASYNC_POSTWRITE);
bus_dmamap_unload(sc->sis_tx_tag, txd->tx_dmamap);
m_freem(txd->tx_m);
txd->tx_m = NULL;
if ((txstat & SIS_CMDSTS_PKT_OK) != 0) {
ifp->if_opackets++;
ifp->if_collisions +=
(txstat & SIS_TXSTAT_COLLCNT) >> 16;
} else {
ifp->if_oerrors++;
if (txstat & SIS_TXSTAT_EXCESSCOLLS)
ifp->if_collisions++;
if (txstat & SIS_TXSTAT_OUTOFWINCOLL)
ifp->if_collisions++;
}
}
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
sc->sis_tx_cnt--;
ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
}
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
sc->sis_tx_cons = cons;
if (sc->sis_tx_cnt == 0)
sc->sis_watchdog_timer = 0;
}
static void
sis_tick(void *xsc)
{
struct sis_softc *sc;
struct mii_data *mii;
struct ifnet *ifp;
sc = xsc;
SIS_LOCK_ASSERT(sc);
ifp = sc->sis_ifp;
mii = device_get_softc(sc->sis_miibus);
mii_tick(mii);
sis_watchdog(sc);
if ((sc->sis_flags & SIS_FLAG_LINK) == 0)
sis_miibus_statchg(sc->sis_dev);
callout_reset(&sc->sis_stat_ch, hz, sis_tick, sc);
}
Device Polling code for -current. Non-SMP, i386-only, no polling in the idle loop at the moment. To use this code you must compile a kernel with options DEVICE_POLLING and at runtime enable polling with sysctl kern.polling.enable=1 The percentage of CPU reserved to userland can be set with sysctl kern.polling.user_frac=NN (default is 50) while the remainder is used by polling device drivers and netisr's. These are the only two variables that you should need to touch. There are a few more parameters in kern.polling but the default values are adequate for all purposes. See the code in kern_poll.c for more details on them. Polling in the idle loop will be implemented shortly by introducing a kernel thread which does the job. Until then, the amount of CPU dedicated to polling will never exceed (100-user_frac). The equivalent (actually, better) code for -stable is at http://info.iet.unipi.it/~luigi/polling/ and also supports polling in the idle loop. NOTE to Alpha developers: There is really nothing in this code that is i386-specific. If you move the 2 lines supporting the new option from sys/conf/{files,options}.i386 to sys/conf/{files,options} I am pretty sure that this should work on the Alpha as well, just that I do not have a suitable test box to try it. If someone feels like trying it, I would appreciate it. NOTE to other developers: sure some things could be done better, and as always I am open to constructive criticism, which a few of you have already given and I greatly appreciated. However, before proposing radical architectural changes, please take some time to possibly try out this code, or at the very least read the comments in kern_poll.c, especially re. the reason why I am using a soft netisr and cannot (I believe) replace it with a simple timeout. Quick description of files touched by this commit: sys/conf/files.i386 new file kern/kern_poll.c sys/conf/options.i386 new option sys/i386/i386/trap.c poll in trap (disabled by default) sys/kern/kern_clock.c initialization and hardclock hooks. sys/kern/kern_intr.c minor swi_net changes sys/kern/kern_poll.c the bulk of the code. sys/net/if.h new flag sys/net/if_var.h declaration for functions used in device drivers. sys/net/netisr.h NETISR_POLL sys/dev/fxp/if_fxp.c sys/dev/fxp/if_fxpvar.h sys/pci/if_dc.c sys/pci/if_dcreg.h sys/pci/if_sis.c sys/pci/if_sisreg.h device driver modifications
2001-12-14 17:56:12 +00:00
#ifdef DEVICE_POLLING
static poll_handler_t sis_poll;
static int
Device Polling code for -current. Non-SMP, i386-only, no polling in the idle loop at the moment. To use this code you must compile a kernel with options DEVICE_POLLING and at runtime enable polling with sysctl kern.polling.enable=1 The percentage of CPU reserved to userland can be set with sysctl kern.polling.user_frac=NN (default is 50) while the remainder is used by polling device drivers and netisr's. These are the only two variables that you should need to touch. There are a few more parameters in kern.polling but the default values are adequate for all purposes. See the code in kern_poll.c for more details on them. Polling in the idle loop will be implemented shortly by introducing a kernel thread which does the job. Until then, the amount of CPU dedicated to polling will never exceed (100-user_frac). The equivalent (actually, better) code for -stable is at http://info.iet.unipi.it/~luigi/polling/ and also supports polling in the idle loop. NOTE to Alpha developers: There is really nothing in this code that is i386-specific. If you move the 2 lines supporting the new option from sys/conf/{files,options}.i386 to sys/conf/{files,options} I am pretty sure that this should work on the Alpha as well, just that I do not have a suitable test box to try it. If someone feels like trying it, I would appreciate it. NOTE to other developers: sure some things could be done better, and as always I am open to constructive criticism, which a few of you have already given and I greatly appreciated. However, before proposing radical architectural changes, please take some time to possibly try out this code, or at the very least read the comments in kern_poll.c, especially re. the reason why I am using a soft netisr and cannot (I believe) replace it with a simple timeout. Quick description of files touched by this commit: sys/conf/files.i386 new file kern/kern_poll.c sys/conf/options.i386 new option sys/i386/i386/trap.c poll in trap (disabled by default) sys/kern/kern_clock.c initialization and hardclock hooks. sys/kern/kern_intr.c minor swi_net changes sys/kern/kern_poll.c the bulk of the code. sys/net/if.h new flag sys/net/if_var.h declaration for functions used in device drivers. sys/net/netisr.h NETISR_POLL sys/dev/fxp/if_fxp.c sys/dev/fxp/if_fxpvar.h sys/pci/if_dc.c sys/pci/if_dcreg.h sys/pci/if_sis.c sys/pci/if_sisreg.h device driver modifications
2001-12-14 17:56:12 +00:00
sis_poll(struct ifnet *ifp, enum poll_cmd cmd, int count)
{
struct sis_softc *sc = ifp->if_softc;
int rx_npkts = 0;
Device Polling code for -current. Non-SMP, i386-only, no polling in the idle loop at the moment. To use this code you must compile a kernel with options DEVICE_POLLING and at runtime enable polling with sysctl kern.polling.enable=1 The percentage of CPU reserved to userland can be set with sysctl kern.polling.user_frac=NN (default is 50) while the remainder is used by polling device drivers and netisr's. These are the only two variables that you should need to touch. There are a few more parameters in kern.polling but the default values are adequate for all purposes. See the code in kern_poll.c for more details on them. Polling in the idle loop will be implemented shortly by introducing a kernel thread which does the job. Until then, the amount of CPU dedicated to polling will never exceed (100-user_frac). The equivalent (actually, better) code for -stable is at http://info.iet.unipi.it/~luigi/polling/ and also supports polling in the idle loop. NOTE to Alpha developers: There is really nothing in this code that is i386-specific. If you move the 2 lines supporting the new option from sys/conf/{files,options}.i386 to sys/conf/{files,options} I am pretty sure that this should work on the Alpha as well, just that I do not have a suitable test box to try it. If someone feels like trying it, I would appreciate it. NOTE to other developers: sure some things could be done better, and as always I am open to constructive criticism, which a few of you have already given and I greatly appreciated. However, before proposing radical architectural changes, please take some time to possibly try out this code, or at the very least read the comments in kern_poll.c, especially re. the reason why I am using a soft netisr and cannot (I believe) replace it with a simple timeout. Quick description of files touched by this commit: sys/conf/files.i386 new file kern/kern_poll.c sys/conf/options.i386 new option sys/i386/i386/trap.c poll in trap (disabled by default) sys/kern/kern_clock.c initialization and hardclock hooks. sys/kern/kern_intr.c minor swi_net changes sys/kern/kern_poll.c the bulk of the code. sys/net/if.h new flag sys/net/if_var.h declaration for functions used in device drivers. sys/net/netisr.h NETISR_POLL sys/dev/fxp/if_fxp.c sys/dev/fxp/if_fxpvar.h sys/pci/if_dc.c sys/pci/if_dcreg.h sys/pci/if_sis.c sys/pci/if_sisreg.h device driver modifications
2001-12-14 17:56:12 +00:00
SIS_LOCK(sc);
Big polling(4) cleanup. o Axe poll in trap. o Axe IFF_POLLING flag from if_flags. o Rework revision 1.21 (Giant removal), in such a way that poll_mtx is not dropped during call to polling handler. This fixes problem with idle polling. o Make registration and deregistration from polling in a functional way, insted of next tick/interrupt. o Obsolete kern.polling.enable. Polling is turned on/off with ifconfig. Detailed kern_poll.c changes: - Remove polling handler flags, introduced in 1.21. The are not needed now. - Forget and do not check if_flags, if_capenable and if_drv_flags. - Call all registered polling handlers unconditionally. - Do not drop poll_mtx, when entering polling handlers. - In ether_poll() NET_LOCK_GIANT prior to locking poll_mtx. - In netisr_poll() axe the block, where polling code asks drivers to unregister. - In netisr_poll() and ether_poll() do polling always, if any handlers are present. - In ether_poll_[de]register() remove a lot of error hiding code. Assert that arguments are correct, instead. - In ether_poll_[de]register() use standard return values in case of error or success. - Introduce poll_switch() that is a sysctl handler for kern.polling.enable. poll_switch() goes through interface list and enabled/disables polling. A message that kern.polling.enable is deprecated is printed. Detailed driver changes: - On attach driver announces IFCAP_POLLING in if_capabilities, but not in if_capenable. - On detach driver calls ether_poll_deregister() if polling is enabled. - In polling handler driver obtains its lock and checks IFF_DRV_RUNNING flag. If there is no, then unlocks and returns. - In ioctl handler driver checks for IFCAP_POLLING flag requested to be set or cleared. Driver first calls ether_poll_[de]register(), then obtains driver lock and [dis/en]ables interrupts. - In interrupt handler driver checks IFCAP_POLLING flag in if_capenable. If present, then returns.This is important to protect from spurious interrupts. Reviewed by: ru, sam, jhb
2005-10-01 18:56:19 +00:00
if (!(ifp->if_drv_flags & IFF_DRV_RUNNING)) {
SIS_UNLOCK(sc);
return (rx_npkts);
Device Polling code for -current. Non-SMP, i386-only, no polling in the idle loop at the moment. To use this code you must compile a kernel with options DEVICE_POLLING and at runtime enable polling with sysctl kern.polling.enable=1 The percentage of CPU reserved to userland can be set with sysctl kern.polling.user_frac=NN (default is 50) while the remainder is used by polling device drivers and netisr's. These are the only two variables that you should need to touch. There are a few more parameters in kern.polling but the default values are adequate for all purposes. See the code in kern_poll.c for more details on them. Polling in the idle loop will be implemented shortly by introducing a kernel thread which does the job. Until then, the amount of CPU dedicated to polling will never exceed (100-user_frac). The equivalent (actually, better) code for -stable is at http://info.iet.unipi.it/~luigi/polling/ and also supports polling in the idle loop. NOTE to Alpha developers: There is really nothing in this code that is i386-specific. If you move the 2 lines supporting the new option from sys/conf/{files,options}.i386 to sys/conf/{files,options} I am pretty sure that this should work on the Alpha as well, just that I do not have a suitable test box to try it. If someone feels like trying it, I would appreciate it. NOTE to other developers: sure some things could be done better, and as always I am open to constructive criticism, which a few of you have already given and I greatly appreciated. However, before proposing radical architectural changes, please take some time to possibly try out this code, or at the very least read the comments in kern_poll.c, especially re. the reason why I am using a soft netisr and cannot (I believe) replace it with a simple timeout. Quick description of files touched by this commit: sys/conf/files.i386 new file kern/kern_poll.c sys/conf/options.i386 new option sys/i386/i386/trap.c poll in trap (disabled by default) sys/kern/kern_clock.c initialization and hardclock hooks. sys/kern/kern_intr.c minor swi_net changes sys/kern/kern_poll.c the bulk of the code. sys/net/if.h new flag sys/net/if_var.h declaration for functions used in device drivers. sys/net/netisr.h NETISR_POLL sys/dev/fxp/if_fxp.c sys/dev/fxp/if_fxpvar.h sys/pci/if_dc.c sys/pci/if_dcreg.h sys/pci/if_sis.c sys/pci/if_sisreg.h device driver modifications
2001-12-14 17:56:12 +00:00
}
/*
* On the sis, reading the status register also clears it.
* So before returning to intr mode we must make sure that all
* possible pending sources of interrupts have been served.
* In practice this means run to completion the *eof routines,
* and then call the interrupt routine
*/
sc->rxcycles = count;
rx_npkts = sis_rxeof(sc);
Device Polling code for -current. Non-SMP, i386-only, no polling in the idle loop at the moment. To use this code you must compile a kernel with options DEVICE_POLLING and at runtime enable polling with sysctl kern.polling.enable=1 The percentage of CPU reserved to userland can be set with sysctl kern.polling.user_frac=NN (default is 50) while the remainder is used by polling device drivers and netisr's. These are the only two variables that you should need to touch. There are a few more parameters in kern.polling but the default values are adequate for all purposes. See the code in kern_poll.c for more details on them. Polling in the idle loop will be implemented shortly by introducing a kernel thread which does the job. Until then, the amount of CPU dedicated to polling will never exceed (100-user_frac). The equivalent (actually, better) code for -stable is at http://info.iet.unipi.it/~luigi/polling/ and also supports polling in the idle loop. NOTE to Alpha developers: There is really nothing in this code that is i386-specific. If you move the 2 lines supporting the new option from sys/conf/{files,options}.i386 to sys/conf/{files,options} I am pretty sure that this should work on the Alpha as well, just that I do not have a suitable test box to try it. If someone feels like trying it, I would appreciate it. NOTE to other developers: sure some things could be done better, and as always I am open to constructive criticism, which a few of you have already given and I greatly appreciated. However, before proposing radical architectural changes, please take some time to possibly try out this code, or at the very least read the comments in kern_poll.c, especially re. the reason why I am using a soft netisr and cannot (I believe) replace it with a simple timeout. Quick description of files touched by this commit: sys/conf/files.i386 new file kern/kern_poll.c sys/conf/options.i386 new option sys/i386/i386/trap.c poll in trap (disabled by default) sys/kern/kern_clock.c initialization and hardclock hooks. sys/kern/kern_intr.c minor swi_net changes sys/kern/kern_poll.c the bulk of the code. sys/net/if.h new flag sys/net/if_var.h declaration for functions used in device drivers. sys/net/netisr.h NETISR_POLL sys/dev/fxp/if_fxp.c sys/dev/fxp/if_fxpvar.h sys/pci/if_dc.c sys/pci/if_dcreg.h sys/pci/if_sis.c sys/pci/if_sisreg.h device driver modifications
2001-12-14 17:56:12 +00:00
sis_txeof(sc);
if (!IFQ_DRV_IS_EMPTY(&ifp->if_snd))
sis_startl(ifp);
Device Polling code for -current. Non-SMP, i386-only, no polling in the idle loop at the moment. To use this code you must compile a kernel with options DEVICE_POLLING and at runtime enable polling with sysctl kern.polling.enable=1 The percentage of CPU reserved to userland can be set with sysctl kern.polling.user_frac=NN (default is 50) while the remainder is used by polling device drivers and netisr's. These are the only two variables that you should need to touch. There are a few more parameters in kern.polling but the default values are adequate for all purposes. See the code in kern_poll.c for more details on them. Polling in the idle loop will be implemented shortly by introducing a kernel thread which does the job. Until then, the amount of CPU dedicated to polling will never exceed (100-user_frac). The equivalent (actually, better) code for -stable is at http://info.iet.unipi.it/~luigi/polling/ and also supports polling in the idle loop. NOTE to Alpha developers: There is really nothing in this code that is i386-specific. If you move the 2 lines supporting the new option from sys/conf/{files,options}.i386 to sys/conf/{files,options} I am pretty sure that this should work on the Alpha as well, just that I do not have a suitable test box to try it. If someone feels like trying it, I would appreciate it. NOTE to other developers: sure some things could be done better, and as always I am open to constructive criticism, which a few of you have already given and I greatly appreciated. However, before proposing radical architectural changes, please take some time to possibly try out this code, or at the very least read the comments in kern_poll.c, especially re. the reason why I am using a soft netisr and cannot (I believe) replace it with a simple timeout. Quick description of files touched by this commit: sys/conf/files.i386 new file kern/kern_poll.c sys/conf/options.i386 new option sys/i386/i386/trap.c poll in trap (disabled by default) sys/kern/kern_clock.c initialization and hardclock hooks. sys/kern/kern_intr.c minor swi_net changes sys/kern/kern_poll.c the bulk of the code. sys/net/if.h new flag sys/net/if_var.h declaration for functions used in device drivers. sys/net/netisr.h NETISR_POLL sys/dev/fxp/if_fxp.c sys/dev/fxp/if_fxpvar.h sys/pci/if_dc.c sys/pci/if_dcreg.h sys/pci/if_sis.c sys/pci/if_sisreg.h device driver modifications
2001-12-14 17:56:12 +00:00
if (sc->rxcycles > 0 || cmd == POLL_AND_CHECK_STATUS) {
2010-09-01 19:53:15 +00:00
uint32_t status;
Device Polling code for -current. Non-SMP, i386-only, no polling in the idle loop at the moment. To use this code you must compile a kernel with options DEVICE_POLLING and at runtime enable polling with sysctl kern.polling.enable=1 The percentage of CPU reserved to userland can be set with sysctl kern.polling.user_frac=NN (default is 50) while the remainder is used by polling device drivers and netisr's. These are the only two variables that you should need to touch. There are a few more parameters in kern.polling but the default values are adequate for all purposes. See the code in kern_poll.c for more details on them. Polling in the idle loop will be implemented shortly by introducing a kernel thread which does the job. Until then, the amount of CPU dedicated to polling will never exceed (100-user_frac). The equivalent (actually, better) code for -stable is at http://info.iet.unipi.it/~luigi/polling/ and also supports polling in the idle loop. NOTE to Alpha developers: There is really nothing in this code that is i386-specific. If you move the 2 lines supporting the new option from sys/conf/{files,options}.i386 to sys/conf/{files,options} I am pretty sure that this should work on the Alpha as well, just that I do not have a suitable test box to try it. If someone feels like trying it, I would appreciate it. NOTE to other developers: sure some things could be done better, and as always I am open to constructive criticism, which a few of you have already given and I greatly appreciated. However, before proposing radical architectural changes, please take some time to possibly try out this code, or at the very least read the comments in kern_poll.c, especially re. the reason why I am using a soft netisr and cannot (I believe) replace it with a simple timeout. Quick description of files touched by this commit: sys/conf/files.i386 new file kern/kern_poll.c sys/conf/options.i386 new option sys/i386/i386/trap.c poll in trap (disabled by default) sys/kern/kern_clock.c initialization and hardclock hooks. sys/kern/kern_intr.c minor swi_net changes sys/kern/kern_poll.c the bulk of the code. sys/net/if.h new flag sys/net/if_var.h declaration for functions used in device drivers. sys/net/netisr.h NETISR_POLL sys/dev/fxp/if_fxp.c sys/dev/fxp/if_fxpvar.h sys/pci/if_dc.c sys/pci/if_dcreg.h sys/pci/if_sis.c sys/pci/if_sisreg.h device driver modifications
2001-12-14 17:56:12 +00:00
/* Reading the ISR register clears all interrupts. */
status = CSR_READ_4(sc, SIS_ISR);
if (status & (SIS_ISR_RX_ERR|SIS_ISR_RX_OFLOW))
ifp->if_ierrors++;
Device Polling code for -current. Non-SMP, i386-only, no polling in the idle loop at the moment. To use this code you must compile a kernel with options DEVICE_POLLING and at runtime enable polling with sysctl kern.polling.enable=1 The percentage of CPU reserved to userland can be set with sysctl kern.polling.user_frac=NN (default is 50) while the remainder is used by polling device drivers and netisr's. These are the only two variables that you should need to touch. There are a few more parameters in kern.polling but the default values are adequate for all purposes. See the code in kern_poll.c for more details on them. Polling in the idle loop will be implemented shortly by introducing a kernel thread which does the job. Until then, the amount of CPU dedicated to polling will never exceed (100-user_frac). The equivalent (actually, better) code for -stable is at http://info.iet.unipi.it/~luigi/polling/ and also supports polling in the idle loop. NOTE to Alpha developers: There is really nothing in this code that is i386-specific. If you move the 2 lines supporting the new option from sys/conf/{files,options}.i386 to sys/conf/{files,options} I am pretty sure that this should work on the Alpha as well, just that I do not have a suitable test box to try it. If someone feels like trying it, I would appreciate it. NOTE to other developers: sure some things could be done better, and as always I am open to constructive criticism, which a few of you have already given and I greatly appreciated. However, before proposing radical architectural changes, please take some time to possibly try out this code, or at the very least read the comments in kern_poll.c, especially re. the reason why I am using a soft netisr and cannot (I believe) replace it with a simple timeout. Quick description of files touched by this commit: sys/conf/files.i386 new file kern/kern_poll.c sys/conf/options.i386 new option sys/i386/i386/trap.c poll in trap (disabled by default) sys/kern/kern_clock.c initialization and hardclock hooks. sys/kern/kern_intr.c minor swi_net changes sys/kern/kern_poll.c the bulk of the code. sys/net/if.h new flag sys/net/if_var.h declaration for functions used in device drivers. sys/net/netisr.h NETISR_POLL sys/dev/fxp/if_fxp.c sys/dev/fxp/if_fxpvar.h sys/pci/if_dc.c sys/pci/if_dcreg.h sys/pci/if_sis.c sys/pci/if_sisreg.h device driver modifications
2001-12-14 17:56:12 +00:00
if (status & (SIS_ISR_RX_IDLE))
SIS_SETBIT(sc, SIS_CSR, SIS_CSR_RX_ENABLE);
if (status & SIS_ISR_SYSERR) {
ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
sis_initl(sc);
Device Polling code for -current. Non-SMP, i386-only, no polling in the idle loop at the moment. To use this code you must compile a kernel with options DEVICE_POLLING and at runtime enable polling with sysctl kern.polling.enable=1 The percentage of CPU reserved to userland can be set with sysctl kern.polling.user_frac=NN (default is 50) while the remainder is used by polling device drivers and netisr's. These are the only two variables that you should need to touch. There are a few more parameters in kern.polling but the default values are adequate for all purposes. See the code in kern_poll.c for more details on them. Polling in the idle loop will be implemented shortly by introducing a kernel thread which does the job. Until then, the amount of CPU dedicated to polling will never exceed (100-user_frac). The equivalent (actually, better) code for -stable is at http://info.iet.unipi.it/~luigi/polling/ and also supports polling in the idle loop. NOTE to Alpha developers: There is really nothing in this code that is i386-specific. If you move the 2 lines supporting the new option from sys/conf/{files,options}.i386 to sys/conf/{files,options} I am pretty sure that this should work on the Alpha as well, just that I do not have a suitable test box to try it. If someone feels like trying it, I would appreciate it. NOTE to other developers: sure some things could be done better, and as always I am open to constructive criticism, which a few of you have already given and I greatly appreciated. However, before proposing radical architectural changes, please take some time to possibly try out this code, or at the very least read the comments in kern_poll.c, especially re. the reason why I am using a soft netisr and cannot (I believe) replace it with a simple timeout. Quick description of files touched by this commit: sys/conf/files.i386 new file kern/kern_poll.c sys/conf/options.i386 new option sys/i386/i386/trap.c poll in trap (disabled by default) sys/kern/kern_clock.c initialization and hardclock hooks. sys/kern/kern_intr.c minor swi_net changes sys/kern/kern_poll.c the bulk of the code. sys/net/if.h new flag sys/net/if_var.h declaration for functions used in device drivers. sys/net/netisr.h NETISR_POLL sys/dev/fxp/if_fxp.c sys/dev/fxp/if_fxpvar.h sys/pci/if_dc.c sys/pci/if_dcreg.h sys/pci/if_sis.c sys/pci/if_sisreg.h device driver modifications
2001-12-14 17:56:12 +00:00
}
}
Big polling(4) cleanup. o Axe poll in trap. o Axe IFF_POLLING flag from if_flags. o Rework revision 1.21 (Giant removal), in such a way that poll_mtx is not dropped during call to polling handler. This fixes problem with idle polling. o Make registration and deregistration from polling in a functional way, insted of next tick/interrupt. o Obsolete kern.polling.enable. Polling is turned on/off with ifconfig. Detailed kern_poll.c changes: - Remove polling handler flags, introduced in 1.21. The are not needed now. - Forget and do not check if_flags, if_capenable and if_drv_flags. - Call all registered polling handlers unconditionally. - Do not drop poll_mtx, when entering polling handlers. - In ether_poll() NET_LOCK_GIANT prior to locking poll_mtx. - In netisr_poll() axe the block, where polling code asks drivers to unregister. - In netisr_poll() and ether_poll() do polling always, if any handlers are present. - In ether_poll_[de]register() remove a lot of error hiding code. Assert that arguments are correct, instead. - In ether_poll_[de]register() use standard return values in case of error or success. - Introduce poll_switch() that is a sysctl handler for kern.polling.enable. poll_switch() goes through interface list and enabled/disables polling. A message that kern.polling.enable is deprecated is printed. Detailed driver changes: - On attach driver announces IFCAP_POLLING in if_capabilities, but not in if_capenable. - On detach driver calls ether_poll_deregister() if polling is enabled. - In polling handler driver obtains its lock and checks IFF_DRV_RUNNING flag. If there is no, then unlocks and returns. - In ioctl handler driver checks for IFCAP_POLLING flag requested to be set or cleared. Driver first calls ether_poll_[de]register(), then obtains driver lock and [dis/en]ables interrupts. - In interrupt handler driver checks IFCAP_POLLING flag in if_capenable. If present, then returns.This is important to protect from spurious interrupts. Reviewed by: ru, sam, jhb
2005-10-01 18:56:19 +00:00
Device Polling code for -current. Non-SMP, i386-only, no polling in the idle loop at the moment. To use this code you must compile a kernel with options DEVICE_POLLING and at runtime enable polling with sysctl kern.polling.enable=1 The percentage of CPU reserved to userland can be set with sysctl kern.polling.user_frac=NN (default is 50) while the remainder is used by polling device drivers and netisr's. These are the only two variables that you should need to touch. There are a few more parameters in kern.polling but the default values are adequate for all purposes. See the code in kern_poll.c for more details on them. Polling in the idle loop will be implemented shortly by introducing a kernel thread which does the job. Until then, the amount of CPU dedicated to polling will never exceed (100-user_frac). The equivalent (actually, better) code for -stable is at http://info.iet.unipi.it/~luigi/polling/ and also supports polling in the idle loop. NOTE to Alpha developers: There is really nothing in this code that is i386-specific. If you move the 2 lines supporting the new option from sys/conf/{files,options}.i386 to sys/conf/{files,options} I am pretty sure that this should work on the Alpha as well, just that I do not have a suitable test box to try it. If someone feels like trying it, I would appreciate it. NOTE to other developers: sure some things could be done better, and as always I am open to constructive criticism, which a few of you have already given and I greatly appreciated. However, before proposing radical architectural changes, please take some time to possibly try out this code, or at the very least read the comments in kern_poll.c, especially re. the reason why I am using a soft netisr and cannot (I believe) replace it with a simple timeout. Quick description of files touched by this commit: sys/conf/files.i386 new file kern/kern_poll.c sys/conf/options.i386 new option sys/i386/i386/trap.c poll in trap (disabled by default) sys/kern/kern_clock.c initialization and hardclock hooks. sys/kern/kern_intr.c minor swi_net changes sys/kern/kern_poll.c the bulk of the code. sys/net/if.h new flag sys/net/if_var.h declaration for functions used in device drivers. sys/net/netisr.h NETISR_POLL sys/dev/fxp/if_fxp.c sys/dev/fxp/if_fxpvar.h sys/pci/if_dc.c sys/pci/if_dcreg.h sys/pci/if_sis.c sys/pci/if_sisreg.h device driver modifications
2001-12-14 17:56:12 +00:00
SIS_UNLOCK(sc);
return (rx_npkts);
Device Polling code for -current. Non-SMP, i386-only, no polling in the idle loop at the moment. To use this code you must compile a kernel with options DEVICE_POLLING and at runtime enable polling with sysctl kern.polling.enable=1 The percentage of CPU reserved to userland can be set with sysctl kern.polling.user_frac=NN (default is 50) while the remainder is used by polling device drivers and netisr's. These are the only two variables that you should need to touch. There are a few more parameters in kern.polling but the default values are adequate for all purposes. See the code in kern_poll.c for more details on them. Polling in the idle loop will be implemented shortly by introducing a kernel thread which does the job. Until then, the amount of CPU dedicated to polling will never exceed (100-user_frac). The equivalent (actually, better) code for -stable is at http://info.iet.unipi.it/~luigi/polling/ and also supports polling in the idle loop. NOTE to Alpha developers: There is really nothing in this code that is i386-specific. If you move the 2 lines supporting the new option from sys/conf/{files,options}.i386 to sys/conf/{files,options} I am pretty sure that this should work on the Alpha as well, just that I do not have a suitable test box to try it. If someone feels like trying it, I would appreciate it. NOTE to other developers: sure some things could be done better, and as always I am open to constructive criticism, which a few of you have already given and I greatly appreciated. However, before proposing radical architectural changes, please take some time to possibly try out this code, or at the very least read the comments in kern_poll.c, especially re. the reason why I am using a soft netisr and cannot (I believe) replace it with a simple timeout. Quick description of files touched by this commit: sys/conf/files.i386 new file kern/kern_poll.c sys/conf/options.i386 new option sys/i386/i386/trap.c poll in trap (disabled by default) sys/kern/kern_clock.c initialization and hardclock hooks. sys/kern/kern_intr.c minor swi_net changes sys/kern/kern_poll.c the bulk of the code. sys/net/if.h new flag sys/net/if_var.h declaration for functions used in device drivers. sys/net/netisr.h NETISR_POLL sys/dev/fxp/if_fxp.c sys/dev/fxp/if_fxpvar.h sys/pci/if_dc.c sys/pci/if_dcreg.h sys/pci/if_sis.c sys/pci/if_sisreg.h device driver modifications
2001-12-14 17:56:12 +00:00
}
#endif /* DEVICE_POLLING */
static void
sis_intr(void *arg)
{
struct sis_softc *sc;
struct ifnet *ifp;
2010-09-01 19:53:15 +00:00
uint32_t status;
sc = arg;
ifp = sc->sis_ifp;
SIS_LOCK(sc);
Device Polling code for -current. Non-SMP, i386-only, no polling in the idle loop at the moment. To use this code you must compile a kernel with options DEVICE_POLLING and at runtime enable polling with sysctl kern.polling.enable=1 The percentage of CPU reserved to userland can be set with sysctl kern.polling.user_frac=NN (default is 50) while the remainder is used by polling device drivers and netisr's. These are the only two variables that you should need to touch. There are a few more parameters in kern.polling but the default values are adequate for all purposes. See the code in kern_poll.c for more details on them. Polling in the idle loop will be implemented shortly by introducing a kernel thread which does the job. Until then, the amount of CPU dedicated to polling will never exceed (100-user_frac). The equivalent (actually, better) code for -stable is at http://info.iet.unipi.it/~luigi/polling/ and also supports polling in the idle loop. NOTE to Alpha developers: There is really nothing in this code that is i386-specific. If you move the 2 lines supporting the new option from sys/conf/{files,options}.i386 to sys/conf/{files,options} I am pretty sure that this should work on the Alpha as well, just that I do not have a suitable test box to try it. If someone feels like trying it, I would appreciate it. NOTE to other developers: sure some things could be done better, and as always I am open to constructive criticism, which a few of you have already given and I greatly appreciated. However, before proposing radical architectural changes, please take some time to possibly try out this code, or at the very least read the comments in kern_poll.c, especially re. the reason why I am using a soft netisr and cannot (I believe) replace it with a simple timeout. Quick description of files touched by this commit: sys/conf/files.i386 new file kern/kern_poll.c sys/conf/options.i386 new option sys/i386/i386/trap.c poll in trap (disabled by default) sys/kern/kern_clock.c initialization and hardclock hooks. sys/kern/kern_intr.c minor swi_net changes sys/kern/kern_poll.c the bulk of the code. sys/net/if.h new flag sys/net/if_var.h declaration for functions used in device drivers. sys/net/netisr.h NETISR_POLL sys/dev/fxp/if_fxp.c sys/dev/fxp/if_fxpvar.h sys/pci/if_dc.c sys/pci/if_dcreg.h sys/pci/if_sis.c sys/pci/if_sisreg.h device driver modifications
2001-12-14 17:56:12 +00:00
#ifdef DEVICE_POLLING
Big polling(4) cleanup. o Axe poll in trap. o Axe IFF_POLLING flag from if_flags. o Rework revision 1.21 (Giant removal), in such a way that poll_mtx is not dropped during call to polling handler. This fixes problem with idle polling. o Make registration and deregistration from polling in a functional way, insted of next tick/interrupt. o Obsolete kern.polling.enable. Polling is turned on/off with ifconfig. Detailed kern_poll.c changes: - Remove polling handler flags, introduced in 1.21. The are not needed now. - Forget and do not check if_flags, if_capenable and if_drv_flags. - Call all registered polling handlers unconditionally. - Do not drop poll_mtx, when entering polling handlers. - In ether_poll() NET_LOCK_GIANT prior to locking poll_mtx. - In netisr_poll() axe the block, where polling code asks drivers to unregister. - In netisr_poll() and ether_poll() do polling always, if any handlers are present. - In ether_poll_[de]register() remove a lot of error hiding code. Assert that arguments are correct, instead. - In ether_poll_[de]register() use standard return values in case of error or success. - Introduce poll_switch() that is a sysctl handler for kern.polling.enable. poll_switch() goes through interface list and enabled/disables polling. A message that kern.polling.enable is deprecated is printed. Detailed driver changes: - On attach driver announces IFCAP_POLLING in if_capabilities, but not in if_capenable. - On detach driver calls ether_poll_deregister() if polling is enabled. - In polling handler driver obtains its lock and checks IFF_DRV_RUNNING flag. If there is no, then unlocks and returns. - In ioctl handler driver checks for IFCAP_POLLING flag requested to be set or cleared. Driver first calls ether_poll_[de]register(), then obtains driver lock and [dis/en]ables interrupts. - In interrupt handler driver checks IFCAP_POLLING flag in if_capenable. If present, then returns.This is important to protect from spurious interrupts. Reviewed by: ru, sam, jhb
2005-10-01 18:56:19 +00:00
if (ifp->if_capenable & IFCAP_POLLING) {
SIS_UNLOCK(sc);
return;
Device Polling code for -current. Non-SMP, i386-only, no polling in the idle loop at the moment. To use this code you must compile a kernel with options DEVICE_POLLING and at runtime enable polling with sysctl kern.polling.enable=1 The percentage of CPU reserved to userland can be set with sysctl kern.polling.user_frac=NN (default is 50) while the remainder is used by polling device drivers and netisr's. These are the only two variables that you should need to touch. There are a few more parameters in kern.polling but the default values are adequate for all purposes. See the code in kern_poll.c for more details on them. Polling in the idle loop will be implemented shortly by introducing a kernel thread which does the job. Until then, the amount of CPU dedicated to polling will never exceed (100-user_frac). The equivalent (actually, better) code for -stable is at http://info.iet.unipi.it/~luigi/polling/ and also supports polling in the idle loop. NOTE to Alpha developers: There is really nothing in this code that is i386-specific. If you move the 2 lines supporting the new option from sys/conf/{files,options}.i386 to sys/conf/{files,options} I am pretty sure that this should work on the Alpha as well, just that I do not have a suitable test box to try it. If someone feels like trying it, I would appreciate it. NOTE to other developers: sure some things could be done better, and as always I am open to constructive criticism, which a few of you have already given and I greatly appreciated. However, before proposing radical architectural changes, please take some time to possibly try out this code, or at the very least read the comments in kern_poll.c, especially re. the reason why I am using a soft netisr and cannot (I believe) replace it with a simple timeout. Quick description of files touched by this commit: sys/conf/files.i386 new file kern/kern_poll.c sys/conf/options.i386 new option sys/i386/i386/trap.c poll in trap (disabled by default) sys/kern/kern_clock.c initialization and hardclock hooks. sys/kern/kern_intr.c minor swi_net changes sys/kern/kern_poll.c the bulk of the code. sys/net/if.h new flag sys/net/if_var.h declaration for functions used in device drivers. sys/net/netisr.h NETISR_POLL sys/dev/fxp/if_fxp.c sys/dev/fxp/if_fxpvar.h sys/pci/if_dc.c sys/pci/if_dcreg.h sys/pci/if_sis.c sys/pci/if_sisreg.h device driver modifications
2001-12-14 17:56:12 +00:00
}
Big polling(4) cleanup. o Axe poll in trap. o Axe IFF_POLLING flag from if_flags. o Rework revision 1.21 (Giant removal), in such a way that poll_mtx is not dropped during call to polling handler. This fixes problem with idle polling. o Make registration and deregistration from polling in a functional way, insted of next tick/interrupt. o Obsolete kern.polling.enable. Polling is turned on/off with ifconfig. Detailed kern_poll.c changes: - Remove polling handler flags, introduced in 1.21. The are not needed now. - Forget and do not check if_flags, if_capenable and if_drv_flags. - Call all registered polling handlers unconditionally. - Do not drop poll_mtx, when entering polling handlers. - In ether_poll() NET_LOCK_GIANT prior to locking poll_mtx. - In netisr_poll() axe the block, where polling code asks drivers to unregister. - In netisr_poll() and ether_poll() do polling always, if any handlers are present. - In ether_poll_[de]register() remove a lot of error hiding code. Assert that arguments are correct, instead. - In ether_poll_[de]register() use standard return values in case of error or success. - Introduce poll_switch() that is a sysctl handler for kern.polling.enable. poll_switch() goes through interface list and enabled/disables polling. A message that kern.polling.enable is deprecated is printed. Detailed driver changes: - On attach driver announces IFCAP_POLLING in if_capabilities, but not in if_capenable. - On detach driver calls ether_poll_deregister() if polling is enabled. - In polling handler driver obtains its lock and checks IFF_DRV_RUNNING flag. If there is no, then unlocks and returns. - In ioctl handler driver checks for IFCAP_POLLING flag requested to be set or cleared. Driver first calls ether_poll_[de]register(), then obtains driver lock and [dis/en]ables interrupts. - In interrupt handler driver checks IFCAP_POLLING flag in if_capenable. If present, then returns.This is important to protect from spurious interrupts. Reviewed by: ru, sam, jhb
2005-10-01 18:56:19 +00:00
#endif
Device Polling code for -current. Non-SMP, i386-only, no polling in the idle loop at the moment. To use this code you must compile a kernel with options DEVICE_POLLING and at runtime enable polling with sysctl kern.polling.enable=1 The percentage of CPU reserved to userland can be set with sysctl kern.polling.user_frac=NN (default is 50) while the remainder is used by polling device drivers and netisr's. These are the only two variables that you should need to touch. There are a few more parameters in kern.polling but the default values are adequate for all purposes. See the code in kern_poll.c for more details on them. Polling in the idle loop will be implemented shortly by introducing a kernel thread which does the job. Until then, the amount of CPU dedicated to polling will never exceed (100-user_frac). The equivalent (actually, better) code for -stable is at http://info.iet.unipi.it/~luigi/polling/ and also supports polling in the idle loop. NOTE to Alpha developers: There is really nothing in this code that is i386-specific. If you move the 2 lines supporting the new option from sys/conf/{files,options}.i386 to sys/conf/{files,options} I am pretty sure that this should work on the Alpha as well, just that I do not have a suitable test box to try it. If someone feels like trying it, I would appreciate it. NOTE to other developers: sure some things could be done better, and as always I am open to constructive criticism, which a few of you have already given and I greatly appreciated. However, before proposing radical architectural changes, please take some time to possibly try out this code, or at the very least read the comments in kern_poll.c, especially re. the reason why I am using a soft netisr and cannot (I believe) replace it with a simple timeout. Quick description of files touched by this commit: sys/conf/files.i386 new file kern/kern_poll.c sys/conf/options.i386 new option sys/i386/i386/trap.c poll in trap (disabled by default) sys/kern/kern_clock.c initialization and hardclock hooks. sys/kern/kern_intr.c minor swi_net changes sys/kern/kern_poll.c the bulk of the code. sys/net/if.h new flag sys/net/if_var.h declaration for functions used in device drivers. sys/net/netisr.h NETISR_POLL sys/dev/fxp/if_fxp.c sys/dev/fxp/if_fxpvar.h sys/pci/if_dc.c sys/pci/if_dcreg.h sys/pci/if_sis.c sys/pci/if_sisreg.h device driver modifications
2001-12-14 17:56:12 +00:00
/* Reading the ISR register clears all interrupts. */
status = CSR_READ_4(sc, SIS_ISR);
if ((status & SIS_INTRS) == 0) {
/* Not ours. */
SIS_UNLOCK(sc);
}
/* Disable interrupts. */
CSR_WRITE_4(sc, SIS_IER, 0);
for (;(status & SIS_INTRS) != 0;) {
if (status &
(SIS_ISR_TX_DESC_OK | SIS_ISR_TX_ERR |
SIS_ISR_TX_OK | SIS_ISR_TX_IDLE) )
sis_txeof(sc);
if (status & (SIS_ISR_RX_DESC_OK | SIS_ISR_RX_OK |
SIS_ISR_RX_ERR | SIS_ISR_RX_IDLE))
sis_rxeof(sc);
if (status & SIS_ISR_RX_OFLOW)
ifp->if_ierrors++;
if (status & (SIS_ISR_RX_IDLE))
SIS_SETBIT(sc, SIS_CSR, SIS_CSR_RX_ENABLE);
if (status & SIS_ISR_SYSERR) {
ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
sis_initl(sc);
SIS_UNLOCK(sc);
return;
}
status = CSR_READ_4(sc, SIS_ISR);
}
/* Re-enable interrupts. */
CSR_WRITE_4(sc, SIS_IER, 1);
if (!IFQ_DRV_IS_EMPTY(&ifp->if_snd))
sis_startl(ifp);
SIS_UNLOCK(sc);
}
/*
* Encapsulate an mbuf chain in a descriptor by coupling the mbuf data
* pointers to the fragment pointers.
*/
static int
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
sis_encap(struct sis_softc *sc, struct mbuf **m_head)
{
struct mbuf *m;
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
struct sis_txdesc *txd;
struct sis_desc *f;
bus_dma_segment_t segs[SIS_MAXTXSEGS];
bus_dmamap_t map;
int error, i, frag, nsegs, prod;
int padlen;
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
prod = sc->sis_tx_prod;
txd = &sc->sis_txdesc[prod];
if ((sc->sis_flags & SIS_FLAG_MANUAL_PAD) != 0 &&
(*m_head)->m_pkthdr.len < SIS_MIN_FRAMELEN) {
m = *m_head;
padlen = SIS_MIN_FRAMELEN - m->m_pkthdr.len;
if (M_WRITABLE(m) == 0) {
/* Get a writable copy. */
m = m_dup(*m_head, M_DONTWAIT);
m_freem(*m_head);
if (m == NULL) {
*m_head = NULL;
return (ENOBUFS);
}
*m_head = m;
}
if (m->m_next != NULL || M_TRAILINGSPACE(m) < padlen) {
m = m_defrag(m, M_DONTWAIT);
if (m == NULL) {
m_freem(*m_head);
*m_head = NULL;
return (ENOBUFS);
}
}
/*
* Manually pad short frames, and zero the pad space
* to avoid leaking data.
*/
bzero(mtod(m, char *) + m->m_pkthdr.len, padlen);
m->m_pkthdr.len += padlen;
m->m_len = m->m_pkthdr.len;
*m_head = m;
}
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
error = bus_dmamap_load_mbuf_sg(sc->sis_tx_tag, txd->tx_dmamap,
*m_head, segs, &nsegs, 0);
if (error == EFBIG) {
m = m_collapse(*m_head, M_DONTWAIT, SIS_MAXTXSEGS);
if (m == NULL) {
m_freem(*m_head);
*m_head = NULL;
return (ENOBUFS);
}
*m_head = m;
error = bus_dmamap_load_mbuf_sg(sc->sis_tx_tag, txd->tx_dmamap,
*m_head, segs, &nsegs, 0);
if (error != 0) {
m_freem(*m_head);
*m_head = NULL;
return (error);
}
} else if (error != 0)
return (error);
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
/* Check for descriptor overruns. */
if (sc->sis_tx_cnt + nsegs > SIS_TX_LIST_CNT - 1) {
bus_dmamap_unload(sc->sis_tx_tag, txd->tx_dmamap);
return (ENOBUFS);
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
}
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
bus_dmamap_sync(sc->sis_tx_tag, txd->tx_dmamap, BUS_DMASYNC_PREWRITE);
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
frag = prod;
for (i = 0; i < nsegs; i++) {
f = &sc->sis_tx_list[prod];
if (i == 0)
f->sis_cmdsts = htole32(segs[i].ds_len |
SIS_CMDSTS_MORE);
else
f->sis_cmdsts = htole32(segs[i].ds_len |
SIS_CMDSTS_OWN | SIS_CMDSTS_MORE);
f->sis_ptr = htole32(SIS_ADDR_LO(segs[i].ds_addr));
SIS_INC(prod, SIS_TX_LIST_CNT);
sc->sis_tx_cnt++;
}
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
/* Update producer index. */
sc->sis_tx_prod = prod;
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
/* Remove MORE flag on the last descriptor. */
prod = (prod - 1) & (SIS_TX_LIST_CNT - 1);
f = &sc->sis_tx_list[prod];
f->sis_cmdsts &= ~htole32(SIS_CMDSTS_MORE);
/* Lastly transfer ownership of packet to the controller. */
f = &sc->sis_tx_list[frag];
f->sis_cmdsts |= htole32(SIS_CMDSTS_OWN);
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
/* Swap the last and the first dmamaps. */
map = txd->tx_dmamap;
txd->tx_dmamap = sc->sis_txdesc[prod].tx_dmamap;
sc->sis_txdesc[prod].tx_dmamap = map;
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
txd->tx_m = *m_head;
2010-09-01 18:39:35 +00:00
return (0);
}
static void
sis_start(struct ifnet *ifp)
{
struct sis_softc *sc;
sc = ifp->if_softc;
SIS_LOCK(sc);
sis_startl(ifp);
SIS_UNLOCK(sc);
}
static void
sis_startl(struct ifnet *ifp)
{
struct sis_softc *sc;
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
struct mbuf *m_head;
int queued;
sc = ifp->if_softc;
SIS_LOCK_ASSERT(sc);
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
if ((ifp->if_drv_flags & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
IFF_DRV_RUNNING || (sc->sis_flags & SIS_FLAG_LINK) == 0)
return;
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
for (queued = 0; !IFQ_DRV_IS_EMPTY(&ifp->if_snd) &&
sc->sis_tx_cnt < SIS_TX_LIST_CNT - 4;) {
IFQ_DRV_DEQUEUE(&ifp->if_snd, m_head);
if (m_head == NULL)
break;
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
if (sis_encap(sc, &m_head) != 0) {
if (m_head == NULL)
break;
IFQ_DRV_PREPEND(&ifp->if_snd, m_head);
ifp->if_drv_flags |= IFF_DRV_OACTIVE;
break;
}
queued++;
/*
* If there's a BPF listener, bounce a copy of this frame
* to him.
*/
BPF_MTAP(ifp, m_head);
}
if (queued) {
/* Transmit */
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
bus_dmamap_sync(sc->sis_tx_list_tag, sc->sis_tx_list_map,
BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
SIS_SETBIT(sc, SIS_CSR, SIS_CSR_TX_ENABLE);
/*
* Set a timeout in case the chip goes out to lunch.
*/
sc->sis_watchdog_timer = 5;
}
}
static void
sis_init(void *xsc)
{
struct sis_softc *sc = xsc;
SIS_LOCK(sc);
sis_initl(sc);
SIS_UNLOCK(sc);
}
static void
sis_initl(struct sis_softc *sc)
{
struct ifnet *ifp = sc->sis_ifp;
struct mii_data *mii;
uint8_t *eaddr;
SIS_LOCK_ASSERT(sc);
if ((ifp->if_drv_flags & IFF_DRV_RUNNING) != 0)
return;
/*
* Cancel pending I/O and free all RX/TX buffers.
*/
sis_stop(sc);
/*
* Reset the chip to a known state.
*/
sis_reset(sc);
#ifdef notyet
if (sc->sis_type == SIS_TYPE_83815 && sc->sis_srr >= NS_SRR_16A) {
/*
* Configure 400usec of interrupt holdoff. This is based
* on emperical tests on a Soekris 4801.
*/
CSR_WRITE_4(sc, NS_IHR, 0x100 | 4);
}
#endif
mii = device_get_softc(sc->sis_miibus);
/* Set MAC address */
eaddr = IF_LLADDR(sc->sis_ifp);
if (sc->sis_type == SIS_TYPE_83815) {
CSR_WRITE_4(sc, SIS_RXFILT_CTL, NS_FILTADDR_PAR0);
CSR_WRITE_4(sc, SIS_RXFILT_DATA, eaddr[0] | eaddr[1] << 8);
CSR_WRITE_4(sc, SIS_RXFILT_CTL, NS_FILTADDR_PAR1);
CSR_WRITE_4(sc, SIS_RXFILT_DATA, eaddr[2] | eaddr[3] << 8);
CSR_WRITE_4(sc, SIS_RXFILT_CTL, NS_FILTADDR_PAR2);
CSR_WRITE_4(sc, SIS_RXFILT_DATA, eaddr[4] | eaddr[5] << 8);
} else {
CSR_WRITE_4(sc, SIS_RXFILT_CTL, SIS_FILTADDR_PAR0);
CSR_WRITE_4(sc, SIS_RXFILT_DATA, eaddr[0] | eaddr[1] << 8);
CSR_WRITE_4(sc, SIS_RXFILT_CTL, SIS_FILTADDR_PAR1);
CSR_WRITE_4(sc, SIS_RXFILT_DATA, eaddr[2] | eaddr[3] << 8);
CSR_WRITE_4(sc, SIS_RXFILT_CTL, SIS_FILTADDR_PAR2);
CSR_WRITE_4(sc, SIS_RXFILT_DATA, eaddr[4] | eaddr[5] << 8);
}
/* Init circular TX/RX lists. */
if (sis_ring_init(sc) != 0) {
device_printf(sc->sis_dev,
"initialization failed: no memory for rx buffers\n");
sis_stop(sc);
return;
}
if (sc->sis_type == SIS_TYPE_83815 || sc->sis_type == SIS_TYPE_83816) {
if (sc->sis_manual_pad != 0)
sc->sis_flags |= SIS_FLAG_MANUAL_PAD;
else
sc->sis_flags &= ~SIS_FLAG_MANUAL_PAD;
}
/*
* Short Cable Receive Errors (MP21.E)
* also: Page 78 of the DP83815 data sheet (september 2002 version)
* recommends the following register settings "for optimum
* performance." for rev 15C. Set this also for 15D parts as
* they require it in practice.
*/
if (sc->sis_type == SIS_TYPE_83815 && sc->sis_srr <= NS_SRR_15D) {
CSR_WRITE_4(sc, NS_PHY_PAGE, 0x0001);
CSR_WRITE_4(sc, NS_PHY_CR, 0x189C);
/* set val for c2 */
CSR_WRITE_4(sc, NS_PHY_TDATA, 0x0000);
/* load/kill c2 */
CSR_WRITE_4(sc, NS_PHY_DSPCFG, 0x5040);
/* rais SD off, from 4 to c */
CSR_WRITE_4(sc, NS_PHY_SDCFG, 0x008C);
CSR_WRITE_4(sc, NS_PHY_PAGE, 0);
}
/*
* For the NatSemi chip, we have to explicitly enable the
* reception of ARP frames, as well as turn on the 'perfect
* match' filter where we store the station address, otherwise
* we won't receive unicasts meant for this host.
*/
if (sc->sis_type == SIS_TYPE_83815) {
SIS_SETBIT(sc, SIS_RXFILT_CTL, NS_RXFILTCTL_ARP);
SIS_SETBIT(sc, SIS_RXFILT_CTL, NS_RXFILTCTL_PERFECT);
}
/* If we want promiscuous mode, set the allframes bit. */
if (ifp->if_flags & IFF_PROMISC) {
SIS_SETBIT(sc, SIS_RXFILT_CTL, SIS_RXFILTCTL_ALLPHYS);
} else {
SIS_CLRBIT(sc, SIS_RXFILT_CTL, SIS_RXFILTCTL_ALLPHYS);
}
/*
* Set the capture broadcast bit to capture broadcast frames.
*/
if (ifp->if_flags & IFF_BROADCAST) {
SIS_SETBIT(sc, SIS_RXFILT_CTL, SIS_RXFILTCTL_BROAD);
} else {
SIS_CLRBIT(sc, SIS_RXFILT_CTL, SIS_RXFILTCTL_BROAD);
}
/*
* Load the multicast filter.
*/
if (sc->sis_type == SIS_TYPE_83815)
sis_setmulti_ns(sc);
else
sis_setmulti_sis(sc);
/* Turn the receive filter on */
SIS_SETBIT(sc, SIS_RXFILT_CTL, SIS_RXFILTCTL_ENABLE);
/*
* Load the address of the RX and TX lists.
*/
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
CSR_WRITE_4(sc, SIS_RX_LISTPTR, SIS_ADDR_LO(sc->sis_rx_paddr));
CSR_WRITE_4(sc, SIS_TX_LISTPTR, SIS_ADDR_LO(sc->sis_tx_paddr));
/* SIS_CFG_EDB_MASTER_EN indicates the EDB bus is used instead of
* the PCI bus. When this bit is set, the Max DMA Burst Size
* for TX/RX DMA should be no larger than 16 double words.
*/
if (CSR_READ_4(sc, SIS_CFG) & SIS_CFG_EDB_MASTER_EN) {
CSR_WRITE_4(sc, SIS_RX_CFG, SIS_RXCFG64);
} else {
CSR_WRITE_4(sc, SIS_RX_CFG, SIS_RXCFG256);
}
2001-12-05 09:34:28 +00:00
/* Accept Long Packets for VLAN support */
SIS_SETBIT(sc, SIS_RX_CFG, SIS_RXCFG_RX_JABBER);
/*
* Assume 100Mbps link, actual MAC configuration is done
* after getting a valid link.
*/
CSR_WRITE_4(sc, SIS_TX_CFG, SIS_TXCFG_100);
/*
* Enable interrupts.
*/
CSR_WRITE_4(sc, SIS_IMR, SIS_INTRS);
Device Polling code for -current. Non-SMP, i386-only, no polling in the idle loop at the moment. To use this code you must compile a kernel with options DEVICE_POLLING and at runtime enable polling with sysctl kern.polling.enable=1 The percentage of CPU reserved to userland can be set with sysctl kern.polling.user_frac=NN (default is 50) while the remainder is used by polling device drivers and netisr's. These are the only two variables that you should need to touch. There are a few more parameters in kern.polling but the default values are adequate for all purposes. See the code in kern_poll.c for more details on them. Polling in the idle loop will be implemented shortly by introducing a kernel thread which does the job. Until then, the amount of CPU dedicated to polling will never exceed (100-user_frac). The equivalent (actually, better) code for -stable is at http://info.iet.unipi.it/~luigi/polling/ and also supports polling in the idle loop. NOTE to Alpha developers: There is really nothing in this code that is i386-specific. If you move the 2 lines supporting the new option from sys/conf/{files,options}.i386 to sys/conf/{files,options} I am pretty sure that this should work on the Alpha as well, just that I do not have a suitable test box to try it. If someone feels like trying it, I would appreciate it. NOTE to other developers: sure some things could be done better, and as always I am open to constructive criticism, which a few of you have already given and I greatly appreciated. However, before proposing radical architectural changes, please take some time to possibly try out this code, or at the very least read the comments in kern_poll.c, especially re. the reason why I am using a soft netisr and cannot (I believe) replace it with a simple timeout. Quick description of files touched by this commit: sys/conf/files.i386 new file kern/kern_poll.c sys/conf/options.i386 new option sys/i386/i386/trap.c poll in trap (disabled by default) sys/kern/kern_clock.c initialization and hardclock hooks. sys/kern/kern_intr.c minor swi_net changes sys/kern/kern_poll.c the bulk of the code. sys/net/if.h new flag sys/net/if_var.h declaration for functions used in device drivers. sys/net/netisr.h NETISR_POLL sys/dev/fxp/if_fxp.c sys/dev/fxp/if_fxpvar.h sys/pci/if_dc.c sys/pci/if_dcreg.h sys/pci/if_sis.c sys/pci/if_sisreg.h device driver modifications
2001-12-14 17:56:12 +00:00
#ifdef DEVICE_POLLING
/*
* ... only enable interrupts if we are not polling, make sure
* they are off otherwise.
*/
Big polling(4) cleanup. o Axe poll in trap. o Axe IFF_POLLING flag from if_flags. o Rework revision 1.21 (Giant removal), in such a way that poll_mtx is not dropped during call to polling handler. This fixes problem with idle polling. o Make registration and deregistration from polling in a functional way, insted of next tick/interrupt. o Obsolete kern.polling.enable. Polling is turned on/off with ifconfig. Detailed kern_poll.c changes: - Remove polling handler flags, introduced in 1.21. The are not needed now. - Forget and do not check if_flags, if_capenable and if_drv_flags. - Call all registered polling handlers unconditionally. - Do not drop poll_mtx, when entering polling handlers. - In ether_poll() NET_LOCK_GIANT prior to locking poll_mtx. - In netisr_poll() axe the block, where polling code asks drivers to unregister. - In netisr_poll() and ether_poll() do polling always, if any handlers are present. - In ether_poll_[de]register() remove a lot of error hiding code. Assert that arguments are correct, instead. - In ether_poll_[de]register() use standard return values in case of error or success. - Introduce poll_switch() that is a sysctl handler for kern.polling.enable. poll_switch() goes through interface list and enabled/disables polling. A message that kern.polling.enable is deprecated is printed. Detailed driver changes: - On attach driver announces IFCAP_POLLING in if_capabilities, but not in if_capenable. - On detach driver calls ether_poll_deregister() if polling is enabled. - In polling handler driver obtains its lock and checks IFF_DRV_RUNNING flag. If there is no, then unlocks and returns. - In ioctl handler driver checks for IFCAP_POLLING flag requested to be set or cleared. Driver first calls ether_poll_[de]register(), then obtains driver lock and [dis/en]ables interrupts. - In interrupt handler driver checks IFCAP_POLLING flag in if_capenable. If present, then returns.This is important to protect from spurious interrupts. Reviewed by: ru, sam, jhb
2005-10-01 18:56:19 +00:00
if (ifp->if_capenable & IFCAP_POLLING)
Device Polling code for -current. Non-SMP, i386-only, no polling in the idle loop at the moment. To use this code you must compile a kernel with options DEVICE_POLLING and at runtime enable polling with sysctl kern.polling.enable=1 The percentage of CPU reserved to userland can be set with sysctl kern.polling.user_frac=NN (default is 50) while the remainder is used by polling device drivers and netisr's. These are the only two variables that you should need to touch. There are a few more parameters in kern.polling but the default values are adequate for all purposes. See the code in kern_poll.c for more details on them. Polling in the idle loop will be implemented shortly by introducing a kernel thread which does the job. Until then, the amount of CPU dedicated to polling will never exceed (100-user_frac). The equivalent (actually, better) code for -stable is at http://info.iet.unipi.it/~luigi/polling/ and also supports polling in the idle loop. NOTE to Alpha developers: There is really nothing in this code that is i386-specific. If you move the 2 lines supporting the new option from sys/conf/{files,options}.i386 to sys/conf/{files,options} I am pretty sure that this should work on the Alpha as well, just that I do not have a suitable test box to try it. If someone feels like trying it, I would appreciate it. NOTE to other developers: sure some things could be done better, and as always I am open to constructive criticism, which a few of you have already given and I greatly appreciated. However, before proposing radical architectural changes, please take some time to possibly try out this code, or at the very least read the comments in kern_poll.c, especially re. the reason why I am using a soft netisr and cannot (I believe) replace it with a simple timeout. Quick description of files touched by this commit: sys/conf/files.i386 new file kern/kern_poll.c sys/conf/options.i386 new option sys/i386/i386/trap.c poll in trap (disabled by default) sys/kern/kern_clock.c initialization and hardclock hooks. sys/kern/kern_intr.c minor swi_net changes sys/kern/kern_poll.c the bulk of the code. sys/net/if.h new flag sys/net/if_var.h declaration for functions used in device drivers. sys/net/netisr.h NETISR_POLL sys/dev/fxp/if_fxp.c sys/dev/fxp/if_fxpvar.h sys/pci/if_dc.c sys/pci/if_dcreg.h sys/pci/if_sis.c sys/pci/if_sisreg.h device driver modifications
2001-12-14 17:56:12 +00:00
CSR_WRITE_4(sc, SIS_IER, 0);
else
Big polling(4) cleanup. o Axe poll in trap. o Axe IFF_POLLING flag from if_flags. o Rework revision 1.21 (Giant removal), in such a way that poll_mtx is not dropped during call to polling handler. This fixes problem with idle polling. o Make registration and deregistration from polling in a functional way, insted of next tick/interrupt. o Obsolete kern.polling.enable. Polling is turned on/off with ifconfig. Detailed kern_poll.c changes: - Remove polling handler flags, introduced in 1.21. The are not needed now. - Forget and do not check if_flags, if_capenable and if_drv_flags. - Call all registered polling handlers unconditionally. - Do not drop poll_mtx, when entering polling handlers. - In ether_poll() NET_LOCK_GIANT prior to locking poll_mtx. - In netisr_poll() axe the block, where polling code asks drivers to unregister. - In netisr_poll() and ether_poll() do polling always, if any handlers are present. - In ether_poll_[de]register() remove a lot of error hiding code. Assert that arguments are correct, instead. - In ether_poll_[de]register() use standard return values in case of error or success. - Introduce poll_switch() that is a sysctl handler for kern.polling.enable. poll_switch() goes through interface list and enabled/disables polling. A message that kern.polling.enable is deprecated is printed. Detailed driver changes: - On attach driver announces IFCAP_POLLING in if_capabilities, but not in if_capenable. - On detach driver calls ether_poll_deregister() if polling is enabled. - In polling handler driver obtains its lock and checks IFF_DRV_RUNNING flag. If there is no, then unlocks and returns. - In ioctl handler driver checks for IFCAP_POLLING flag requested to be set or cleared. Driver first calls ether_poll_[de]register(), then obtains driver lock and [dis/en]ables interrupts. - In interrupt handler driver checks IFCAP_POLLING flag in if_capenable. If present, then returns.This is important to protect from spurious interrupts. Reviewed by: ru, sam, jhb
2005-10-01 18:56:19 +00:00
#endif
CSR_WRITE_4(sc, SIS_IER, 1);
/* Clear MAC disable. */
SIS_CLRBIT(sc, SIS_CSR, SIS_CSR_TX_DISABLE | SIS_CSR_RX_DISABLE);
sc->sis_flags &= ~SIS_FLAG_LINK;
mii_mediachg(mii);
ifp->if_drv_flags |= IFF_DRV_RUNNING;
ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
callout_reset(&sc->sis_stat_ch, hz, sis_tick, sc);
}
/*
* Set media options.
*/
static int
sis_ifmedia_upd(struct ifnet *ifp)
{
struct sis_softc *sc;
struct mii_data *mii;
int error;
sc = ifp->if_softc;
SIS_LOCK(sc);
mii = device_get_softc(sc->sis_miibus);
if (mii->mii_instance) {
struct mii_softc *miisc;
LIST_FOREACH(miisc, &mii->mii_phys, mii_list)
mii_phy_reset(miisc);
}
error = mii_mediachg(mii);
SIS_UNLOCK(sc);
return (error);
}
/*
* Report current media status.
*/
static void
sis_ifmedia_sts(struct ifnet *ifp, struct ifmediareq *ifmr)
{
struct sis_softc *sc;
struct mii_data *mii;
sc = ifp->if_softc;
SIS_LOCK(sc);
mii = device_get_softc(sc->sis_miibus);
mii_pollstat(mii);
SIS_UNLOCK(sc);
ifmr->ifm_active = mii->mii_media_active;
ifmr->ifm_status = mii->mii_media_status;
}
static int
sis_ioctl(struct ifnet *ifp, u_long command, caddr_t data)
{
struct sis_softc *sc = ifp->if_softc;
struct ifreq *ifr = (struct ifreq *) data;
struct mii_data *mii;
int error = 0;
2010-09-01 18:39:35 +00:00
switch (command) {
case SIOCSIFFLAGS:
SIS_LOCK(sc);
if (ifp->if_flags & IFF_UP) {
if ((ifp->if_drv_flags & IFF_DRV_RUNNING) != 0 &&
((ifp->if_flags ^ sc->sis_if_flags) &
(IFF_PROMISC | IFF_ALLMULTI)) != 0) {
if (sc->sis_type == SIS_TYPE_83815)
sis_setmulti_ns(sc);
else
sis_setmulti_sis(sc);
} else
sis_initl(sc);
} else if (ifp->if_drv_flags & IFF_DRV_RUNNING) {
2005-01-05 22:28:24 +00:00
sis_stop(sc);
}
sc->sis_if_flags = ifp->if_flags;
SIS_UNLOCK(sc);
error = 0;
break;
case SIOCADDMULTI:
case SIOCDELMULTI:
SIS_LOCK(sc);
if (sc->sis_type == SIS_TYPE_83815)
sis_setmulti_ns(sc);
else
sis_setmulti_sis(sc);
SIS_UNLOCK(sc);
break;
case SIOCGIFMEDIA:
case SIOCSIFMEDIA:
mii = device_get_softc(sc->sis_miibus);
error = ifmedia_ioctl(ifp, ifr, &mii->mii_media, command);
break;
case SIOCSIFCAP:
Big polling(4) cleanup. o Axe poll in trap. o Axe IFF_POLLING flag from if_flags. o Rework revision 1.21 (Giant removal), in such a way that poll_mtx is not dropped during call to polling handler. This fixes problem with idle polling. o Make registration and deregistration from polling in a functional way, insted of next tick/interrupt. o Obsolete kern.polling.enable. Polling is turned on/off with ifconfig. Detailed kern_poll.c changes: - Remove polling handler flags, introduced in 1.21. The are not needed now. - Forget and do not check if_flags, if_capenable and if_drv_flags. - Call all registered polling handlers unconditionally. - Do not drop poll_mtx, when entering polling handlers. - In ether_poll() NET_LOCK_GIANT prior to locking poll_mtx. - In netisr_poll() axe the block, where polling code asks drivers to unregister. - In netisr_poll() and ether_poll() do polling always, if any handlers are present. - In ether_poll_[de]register() remove a lot of error hiding code. Assert that arguments are correct, instead. - In ether_poll_[de]register() use standard return values in case of error or success. - Introduce poll_switch() that is a sysctl handler for kern.polling.enable. poll_switch() goes through interface list and enabled/disables polling. A message that kern.polling.enable is deprecated is printed. Detailed driver changes: - On attach driver announces IFCAP_POLLING in if_capabilities, but not in if_capenable. - On detach driver calls ether_poll_deregister() if polling is enabled. - In polling handler driver obtains its lock and checks IFF_DRV_RUNNING flag. If there is no, then unlocks and returns. - In ioctl handler driver checks for IFCAP_POLLING flag requested to be set or cleared. Driver first calls ether_poll_[de]register(), then obtains driver lock and [dis/en]ables interrupts. - In interrupt handler driver checks IFCAP_POLLING flag in if_capenable. If present, then returns.This is important to protect from spurious interrupts. Reviewed by: ru, sam, jhb
2005-10-01 18:56:19 +00:00
/* ok, disable interrupts */
#ifdef DEVICE_POLLING
if (ifr->ifr_reqcap & IFCAP_POLLING &&
!(ifp->if_capenable & IFCAP_POLLING)) {
error = ether_poll_register(sis_poll, ifp);
if (error)
2010-09-01 18:39:35 +00:00
return (error);
Big polling(4) cleanup. o Axe poll in trap. o Axe IFF_POLLING flag from if_flags. o Rework revision 1.21 (Giant removal), in such a way that poll_mtx is not dropped during call to polling handler. This fixes problem with idle polling. o Make registration and deregistration from polling in a functional way, insted of next tick/interrupt. o Obsolete kern.polling.enable. Polling is turned on/off with ifconfig. Detailed kern_poll.c changes: - Remove polling handler flags, introduced in 1.21. The are not needed now. - Forget and do not check if_flags, if_capenable and if_drv_flags. - Call all registered polling handlers unconditionally. - Do not drop poll_mtx, when entering polling handlers. - In ether_poll() NET_LOCK_GIANT prior to locking poll_mtx. - In netisr_poll() axe the block, where polling code asks drivers to unregister. - In netisr_poll() and ether_poll() do polling always, if any handlers are present. - In ether_poll_[de]register() remove a lot of error hiding code. Assert that arguments are correct, instead. - In ether_poll_[de]register() use standard return values in case of error or success. - Introduce poll_switch() that is a sysctl handler for kern.polling.enable. poll_switch() goes through interface list and enabled/disables polling. A message that kern.polling.enable is deprecated is printed. Detailed driver changes: - On attach driver announces IFCAP_POLLING in if_capabilities, but not in if_capenable. - On detach driver calls ether_poll_deregister() if polling is enabled. - In polling handler driver obtains its lock and checks IFF_DRV_RUNNING flag. If there is no, then unlocks and returns. - In ioctl handler driver checks for IFCAP_POLLING flag requested to be set or cleared. Driver first calls ether_poll_[de]register(), then obtains driver lock and [dis/en]ables interrupts. - In interrupt handler driver checks IFCAP_POLLING flag in if_capenable. If present, then returns.This is important to protect from spurious interrupts. Reviewed by: ru, sam, jhb
2005-10-01 18:56:19 +00:00
SIS_LOCK(sc);
/* Disable interrupts */
CSR_WRITE_4(sc, SIS_IER, 0);
ifp->if_capenable |= IFCAP_POLLING;
SIS_UNLOCK(sc);
return (error);
Big polling(4) cleanup. o Axe poll in trap. o Axe IFF_POLLING flag from if_flags. o Rework revision 1.21 (Giant removal), in such a way that poll_mtx is not dropped during call to polling handler. This fixes problem with idle polling. o Make registration and deregistration from polling in a functional way, insted of next tick/interrupt. o Obsolete kern.polling.enable. Polling is turned on/off with ifconfig. Detailed kern_poll.c changes: - Remove polling handler flags, introduced in 1.21. The are not needed now. - Forget and do not check if_flags, if_capenable and if_drv_flags. - Call all registered polling handlers unconditionally. - Do not drop poll_mtx, when entering polling handlers. - In ether_poll() NET_LOCK_GIANT prior to locking poll_mtx. - In netisr_poll() axe the block, where polling code asks drivers to unregister. - In netisr_poll() and ether_poll() do polling always, if any handlers are present. - In ether_poll_[de]register() remove a lot of error hiding code. Assert that arguments are correct, instead. - In ether_poll_[de]register() use standard return values in case of error or success. - Introduce poll_switch() that is a sysctl handler for kern.polling.enable. poll_switch() goes through interface list and enabled/disables polling. A message that kern.polling.enable is deprecated is printed. Detailed driver changes: - On attach driver announces IFCAP_POLLING in if_capabilities, but not in if_capenable. - On detach driver calls ether_poll_deregister() if polling is enabled. - In polling handler driver obtains its lock and checks IFF_DRV_RUNNING flag. If there is no, then unlocks and returns. - In ioctl handler driver checks for IFCAP_POLLING flag requested to be set or cleared. Driver first calls ether_poll_[de]register(), then obtains driver lock and [dis/en]ables interrupts. - In interrupt handler driver checks IFCAP_POLLING flag in if_capenable. If present, then returns.This is important to protect from spurious interrupts. Reviewed by: ru, sam, jhb
2005-10-01 18:56:19 +00:00
}
if (!(ifr->ifr_reqcap & IFCAP_POLLING) &&
ifp->if_capenable & IFCAP_POLLING) {
error = ether_poll_deregister(ifp);
/* Enable interrupts. */
SIS_LOCK(sc);
CSR_WRITE_4(sc, SIS_IER, 1);
ifp->if_capenable &= ~IFCAP_POLLING;
SIS_UNLOCK(sc);
return (error);
}
#endif /* DEVICE_POLLING */
break;
default:
error = ether_ioctl(ifp, command, data);
break;
}
2010-09-01 18:39:35 +00:00
return (error);
}
static void
sis_watchdog(struct sis_softc *sc)
{
SIS_LOCK_ASSERT(sc);
if (sc->sis_watchdog_timer == 0 || --sc->sis_watchdog_timer >0)
return;
device_printf(sc->sis_dev, "watchdog timeout\n");
sc->sis_ifp->if_oerrors++;
sc->sis_ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
sis_initl(sc);
if (!IFQ_DRV_IS_EMPTY(&sc->sis_ifp->if_snd))
sis_startl(sc->sis_ifp);
}
/*
* Stop the adapter and free any mbufs allocated to the
* RX and TX lists.
*/
static void
sis_stop(struct sis_softc *sc)
{
struct ifnet *ifp;
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
struct sis_rxdesc *rxd;
struct sis_txdesc *txd;
int i;
SIS_LOCK_ASSERT(sc);
ifp = sc->sis_ifp;
sc->sis_watchdog_timer = 0;
callout_stop(&sc->sis_stat_ch);
ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
CSR_WRITE_4(sc, SIS_IER, 0);
CSR_WRITE_4(sc, SIS_IMR, 0);
CSR_READ_4(sc, SIS_ISR); /* clear any interrupts already pending */
SIS_SETBIT(sc, SIS_CSR, SIS_CSR_TX_DISABLE|SIS_CSR_RX_DISABLE);
DELAY(1000);
CSR_WRITE_4(sc, SIS_TX_LISTPTR, 0);
CSR_WRITE_4(sc, SIS_RX_LISTPTR, 0);
sc->sis_flags &= ~SIS_FLAG_LINK;
/*
* Free data in the RX lists.
*/
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
for (i = 0; i < SIS_RX_LIST_CNT; i++) {
rxd = &sc->sis_rxdesc[i];
if (rxd->rx_m != NULL) {
bus_dmamap_sync(sc->sis_rx_tag, rxd->rx_dmamap,
BUS_DMASYNC_POSTREAD);
bus_dmamap_unload(sc->sis_rx_tag, rxd->rx_dmamap);
m_freem(rxd->rx_m);
rxd->rx_m = NULL;
}
}
/*
* Free the TX list buffers.
*/
bus_dma(9) cleanup. o Enforce TX/RX descriptor ring alignment. NS data sheet says the controller needs 4 bytes alignment but use 16 to cover both SiS and NS controllers. I don't have SiS data sheet so I'm not sure what is alignment restriction of SiS controller but 16 would be enough because it's larger than the size of a TX/RX descriptor. Previously sis(4) ignored the alignment restriction. o Enforce RX buffer alignment, 4. Previously sis(4) ignored RX buffer alignment restriction. o Limit number of TX DMA segment to be used to 16. It seems controller has no restriction on number of DMA segments but using more than 16 looks resource waste. o Collapse long mbuf chains with m_collapse(9) instead of calling expensive m_defrag(9). o TX/RX side bus_dmamap_load_mbuf_sg(9) support and remove unnecessary callbacks. o Initial endianness support. o Prefer local alignment fixup code to m_devget(9). o Pre-allocate TX/RX mbuf DMA maps instead of creating/destroying these maps in fast TX/RX path. On non-x86 architectures, this is very expensive operation and there is no need to do that. o Add missing bus_dmamap_sync(9) in TX/RX path. o watchdog is now unarmed only when there are no pending frames on controller. Previously sis(4) blindly unarmed watchdog without checking the number of queued frames. o For efficiency, loaded DMA map is reused for error frames. o DMA map loading failure is now gracefully handled. Previously sis(4) ignored any DMA map loading errors. o Nuke unused macros which are not appropriate for endianness operation. o Stop embedding driver maintained structures into descriptor rings. Because TX/RX descriptor structures are shared between host and controller, frequent bus_dmamap_sync(9) operations are required in fast path. Embedding driver structures will increase the size of DMA map which in turn will slow down performance.
2010-09-01 19:33:40 +00:00
for (i = 0; i < SIS_TX_LIST_CNT; i++) {
txd = &sc->sis_txdesc[i];
if (txd->tx_m != NULL) {
bus_dmamap_sync(sc->sis_tx_tag, txd->tx_dmamap,
BUS_DMASYNC_POSTWRITE);
bus_dmamap_unload(sc->sis_tx_tag, txd->tx_dmamap);
m_freem(txd->tx_m);
txd->tx_m = NULL;
}
}
}
/*
* Stop all chip I/O so that the kernel's probe routines don't
* get confused by errant DMAs when rebooting.
*/
2009-02-10 23:17:20 +00:00
static int
sis_shutdown(device_t dev)
{
struct sis_softc *sc;
sc = device_get_softc(dev);
SIS_LOCK(sc);
sis_stop(sc);
SIS_UNLOCK(sc);
2009-02-10 23:17:20 +00:00
return (0);
}
static int
sis_suspend(device_t dev)
{
struct sis_softc *sc;
sc = device_get_softc(dev);
SIS_LOCK(sc);
sis_stop(sc);
SIS_UNLOCK(sc);
return (0);
}
static int
sis_resume(device_t dev)
{
struct sis_softc *sc;
struct ifnet *ifp;
sc = device_get_softc(dev);
SIS_LOCK(sc);
ifp = sc->sis_ifp;
if ((ifp->if_flags & IFF_UP) != 0) {
ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
sis_initl(sc);
}
SIS_UNLOCK(sc);
return (0);
}
static void
sis_add_sysctls(struct sis_softc *sc)
{
struct sysctl_ctx_list *ctx;
struct sysctl_oid_list *children;
char tn[32];
int unit;
ctx = device_get_sysctl_ctx(sc->sis_dev);
children = SYSCTL_CHILDREN(device_get_sysctl_tree(sc->sis_dev));
unit = device_get_unit(sc->sis_dev);
/*
* Unlike most other controllers, NS DP83815/DP83816 controllers
* seem to pad with 0xFF when it encounter short frames. According
* to RFC 1042 the pad bytes should be 0x00. Turning this tunable
* on will have driver pad manully but it's disabled by default
* because it will consume extra CPU cycles for short frames.
*/
sc->sis_manual_pad = 0;
snprintf(tn, sizeof(tn), "dev.sis.%d.manual_pad", unit);
TUNABLE_INT_FETCH(tn, &sc->sis_manual_pad);
SYSCTL_ADD_INT(ctx, children, OID_AUTO, "manual_pad",
CTLFLAG_RW, &sc->sis_manual_pad, 0, "Manually pad short frames");
}
static device_method_t sis_methods[] = {
/* Device interface */
DEVMETHOD(device_probe, sis_probe),
DEVMETHOD(device_attach, sis_attach),
DEVMETHOD(device_detach, sis_detach),
DEVMETHOD(device_shutdown, sis_shutdown),
DEVMETHOD(device_suspend, sis_suspend),
DEVMETHOD(device_resume, sis_resume),
/* bus interface */
DEVMETHOD(bus_print_child, bus_generic_print_child),
DEVMETHOD(bus_driver_added, bus_generic_driver_added),
/* MII interface */
DEVMETHOD(miibus_readreg, sis_miibus_readreg),
DEVMETHOD(miibus_writereg, sis_miibus_writereg),
DEVMETHOD(miibus_statchg, sis_miibus_statchg),
{ 0, 0 }
};
static driver_t sis_driver = {
"sis",
sis_methods,
sizeof(struct sis_softc)
};
static devclass_t sis_devclass;
DRIVER_MODULE(sis, pci, sis_driver, sis_devclass, 0, 0);
DRIVER_MODULE(miibus, sis, miibus_driver, miibus_devclass, 0, 0);