linux-sg2042/drivers/net/mace.c

1022 lines
27 KiB
C
Raw Normal View History

/*
* Network device driver for the MACE ethernet controller on
* Apple Powermacs. Assumes it's under a DBDMA controller.
*
* Copyright (C) 1996 Paul Mackerras.
*/
#include <linux/module.h>
#include <linux/kernel.h>
#include <linux/netdevice.h>
#include <linux/etherdevice.h>
#include <linux/delay.h>
#include <linux/string.h>
#include <linux/timer.h>
#include <linux/init.h>
#include <linux/crc32.h>
#include <linux/spinlock.h>
#include <linux/bitrev.h>
#include <asm/prom.h>
#include <asm/dbdma.h>
#include <asm/io.h>
#include <asm/pgtable.h>
#include <asm/macio.h>
#include "mace.h"
static int port_aaui = -1;
#define N_RX_RING 8
#define N_TX_RING 6
#define MAX_TX_ACTIVE 1
#define NCMDS_TX 1 /* dma commands per element in tx ring */
#define RX_BUFLEN (ETH_FRAME_LEN + 8)
#define TX_TIMEOUT HZ /* 1 second */
/* Chip rev needs workaround on HW & multicast addr change */
#define BROKEN_ADDRCHG_REV 0x0941
/* Bits in transmit DMA status */
#define TX_DMA_ERR 0x80
struct mace_data {
volatile struct mace __iomem *mace;
volatile struct dbdma_regs __iomem *tx_dma;
int tx_dma_intr;
volatile struct dbdma_regs __iomem *rx_dma;
int rx_dma_intr;
volatile struct dbdma_cmd *tx_cmds; /* xmit dma command list */
volatile struct dbdma_cmd *rx_cmds; /* recv dma command list */
struct sk_buff *rx_bufs[N_RX_RING];
int rx_fill;
int rx_empty;
struct sk_buff *tx_bufs[N_TX_RING];
int tx_fill;
int tx_empty;
unsigned char maccc;
unsigned char tx_fullup;
unsigned char tx_active;
unsigned char tx_bad_runt;
struct timer_list tx_timeout;
int timeout_active;
int port_aaui;
int chipid;
struct macio_dev *mdev;
spinlock_t lock;
};
/*
* Number of bytes of private data per MACE: allow enough for
* the rx and tx dma commands plus a branch dma command each,
* and another 16 bytes to allow us to align the dma command
* buffers on a 16 byte boundary.
*/
#define PRIV_BYTES (sizeof(struct mace_data) \
+ (N_RX_RING + NCMDS_TX * N_TX_RING + 3) * sizeof(struct dbdma_cmd))
static int mace_open(struct net_device *dev);
static int mace_close(struct net_device *dev);
static int mace_xmit_start(struct sk_buff *skb, struct net_device *dev);
static void mace_set_multicast(struct net_device *dev);
static void mace_reset(struct net_device *dev);
static int mace_set_address(struct net_device *dev, void *addr);
IRQ: Maintain regs pointer globally rather than passing to IRQ handlers Maintain a per-CPU global "struct pt_regs *" variable which can be used instead of passing regs around manually through all ~1800 interrupt handlers in the Linux kernel. The regs pointer is used in few places, but it potentially costs both stack space and code to pass it around. On the FRV arch, removing the regs parameter from all the genirq function results in a 20% speed up of the IRQ exit path (ie: from leaving timer_interrupt() to leaving do_IRQ()). Where appropriate, an arch may override the generic storage facility and do something different with the variable. On FRV, for instance, the address is maintained in GR28 at all times inside the kernel as part of general exception handling. Having looked over the code, it appears that the parameter may be handed down through up to twenty or so layers of functions. Consider a USB character device attached to a USB hub, attached to a USB controller that posts its interrupts through a cascaded auxiliary interrupt controller. A character device driver may want to pass regs to the sysrq handler through the input layer which adds another few layers of parameter passing. I've build this code with allyesconfig for x86_64 and i386. I've runtested the main part of the code on FRV and i386, though I can't test most of the drivers. I've also done partial conversion for powerpc and MIPS - these at least compile with minimal configurations. This will affect all archs. Mostly the changes should be relatively easy. Take do_IRQ(), store the regs pointer at the beginning, saving the old one: struct pt_regs *old_regs = set_irq_regs(regs); And put the old one back at the end: set_irq_regs(old_regs); Don't pass regs through to generic_handle_irq() or __do_IRQ(). In timer_interrupt(), this sort of change will be necessary: - update_process_times(user_mode(regs)); - profile_tick(CPU_PROFILING, regs); + update_process_times(user_mode(get_irq_regs())); + profile_tick(CPU_PROFILING); I'd like to move update_process_times()'s use of get_irq_regs() into itself, except that i386, alone of the archs, uses something other than user_mode(). Some notes on the interrupt handling in the drivers: (*) input_dev() is now gone entirely. The regs pointer is no longer stored in the input_dev struct. (*) finish_unlinks() in drivers/usb/host/ohci-q.c needs checking. It does something different depending on whether it's been supplied with a regs pointer or not. (*) Various IRQ handler function pointers have been moved to type irq_handler_t. Signed-Off-By: David Howells <dhowells@redhat.com> (cherry picked from 1b16e7ac850969f38b375e511e3fa2f474a33867 commit)
2006-10-05 21:55:46 +08:00
static irqreturn_t mace_interrupt(int irq, void *dev_id);
static irqreturn_t mace_txdma_intr(int irq, void *dev_id);
static irqreturn_t mace_rxdma_intr(int irq, void *dev_id);
static void mace_set_timeout(struct net_device *dev);
static void mace_tx_timeout(unsigned long data);
static inline void dbdma_reset(volatile struct dbdma_regs __iomem *dma);
static inline void mace_clean_rings(struct mace_data *mp);
static void __mace_set_address(struct net_device *dev, void *addr);
/*
* If we can't get a skbuff when we need it, we use this area for DMA.
*/
static unsigned char *dummy_buf;
static int __devinit mace_probe(struct macio_dev *mdev, const struct of_device_id *match)
{
struct device_node *mace = macio_get_of_node(mdev);
struct net_device *dev;
struct mace_data *mp;
const unsigned char *addr;
int j, rev, rc = -EBUSY;
if (macio_resource_count(mdev) != 3 || macio_irq_count(mdev) != 3) {
printk(KERN_ERR "can't use MACE %s: need 3 addrs and 3 irqs\n",
mace->full_name);
return -ENODEV;
}
addr = of_get_property(mace, "mac-address", NULL);
if (addr == NULL) {
addr = of_get_property(mace, "local-mac-address", NULL);
if (addr == NULL) {
printk(KERN_ERR "Can't get mac-address for MACE %s\n",
mace->full_name);
return -ENODEV;
}
}
/*
* lazy allocate the driver-wide dummy buffer. (Note that we
* never have more than one MACE in the system anyway)
*/
if (dummy_buf == NULL) {
dummy_buf = kmalloc(RX_BUFLEN+2, GFP_KERNEL);
if (dummy_buf == NULL) {
printk(KERN_ERR "MACE: couldn't allocate dummy buffer\n");
return -ENOMEM;
}
}
if (macio_request_resources(mdev, "mace")) {
printk(KERN_ERR "MACE: can't request IO resources !\n");
return -EBUSY;
}
dev = alloc_etherdev(PRIV_BYTES);
if (!dev) {
printk(KERN_ERR "MACE: can't allocate ethernet device !\n");
rc = -ENOMEM;
goto err_release;
}
SET_NETDEV_DEV(dev, &mdev->ofdev.dev);
mp = netdev_priv(dev);
mp->mdev = mdev;
macio_set_drvdata(mdev, dev);
dev->base_addr = macio_resource_start(mdev, 0);
mp->mace = ioremap(dev->base_addr, 0x1000);
if (mp->mace == NULL) {
printk(KERN_ERR "MACE: can't map IO resources !\n");
rc = -ENOMEM;
goto err_free;
}
dev->irq = macio_irq(mdev, 0);
rev = addr[0] == 0 && addr[1] == 0xA0;
for (j = 0; j < 6; ++j) {
dev->dev_addr[j] = rev ? bitrev8(addr[j]): addr[j];
}
mp->chipid = (in_8(&mp->mace->chipid_hi) << 8) |
in_8(&mp->mace->chipid_lo);
mp = netdev_priv(dev);
mp->maccc = ENXMT | ENRCV;
mp->tx_dma = ioremap(macio_resource_start(mdev, 1), 0x1000);
if (mp->tx_dma == NULL) {
printk(KERN_ERR "MACE: can't map TX DMA resources !\n");
rc = -ENOMEM;
goto err_unmap_io;
}
mp->tx_dma_intr = macio_irq(mdev, 1);
mp->rx_dma = ioremap(macio_resource_start(mdev, 2), 0x1000);
if (mp->rx_dma == NULL) {
printk(KERN_ERR "MACE: can't map RX DMA resources !\n");
rc = -ENOMEM;
goto err_unmap_tx_dma;
}
mp->rx_dma_intr = macio_irq(mdev, 2);
mp->tx_cmds = (volatile struct dbdma_cmd *) DBDMA_ALIGN(mp + 1);
mp->rx_cmds = mp->tx_cmds + NCMDS_TX * N_TX_RING + 1;
memset((char *) mp->tx_cmds, 0,
(NCMDS_TX*N_TX_RING + N_RX_RING + 2) * sizeof(struct dbdma_cmd));
init_timer(&mp->tx_timeout);
spin_lock_init(&mp->lock);
mp->timeout_active = 0;
if (port_aaui >= 0)
mp->port_aaui = port_aaui;
else {
/* Apple Network Server uses the AAUI port */
if (machine_is_compatible("AAPL,ShinerESB"))
mp->port_aaui = 1;
else {
#ifdef CONFIG_MACE_AAUI_PORT
mp->port_aaui = 1;
#else
mp->port_aaui = 0;
#endif
}
}
dev->open = mace_open;
dev->stop = mace_close;
dev->hard_start_xmit = mace_xmit_start;
dev->set_multicast_list = mace_set_multicast;
dev->set_mac_address = mace_set_address;
/*
* Most of what is below could be moved to mace_open()
*/
mace_reset(dev);
rc = request_irq(dev->irq, mace_interrupt, 0, "MACE", dev);
if (rc) {
printk(KERN_ERR "MACE: can't get irq %d\n", dev->irq);
goto err_unmap_rx_dma;
}
rc = request_irq(mp->tx_dma_intr, mace_txdma_intr, 0, "MACE-txdma", dev);
if (rc) {
2006-07-03 19:36:01 +08:00
printk(KERN_ERR "MACE: can't get irq %d\n", mp->tx_dma_intr);
goto err_free_irq;
}
rc = request_irq(mp->rx_dma_intr, mace_rxdma_intr, 0, "MACE-rxdma", dev);
if (rc) {
2006-07-03 19:36:01 +08:00
printk(KERN_ERR "MACE: can't get irq %d\n", mp->rx_dma_intr);
goto err_free_tx_irq;
}
rc = register_netdev(dev);
if (rc) {
printk(KERN_ERR "MACE: Cannot register net device, aborting.\n");
goto err_free_rx_irq;
}
printk(KERN_INFO "%s: MACE at %pM, chip revision %d.%d\n",
dev->name, dev->dev_addr,
mp->chipid >> 8, mp->chipid & 0xff);
return 0;
err_free_rx_irq:
free_irq(macio_irq(mdev, 2), dev);
err_free_tx_irq:
free_irq(macio_irq(mdev, 1), dev);
err_free_irq:
free_irq(macio_irq(mdev, 0), dev);
err_unmap_rx_dma:
iounmap(mp->rx_dma);
err_unmap_tx_dma:
iounmap(mp->tx_dma);
err_unmap_io:
iounmap(mp->mace);
err_free:
free_netdev(dev);
err_release:
macio_release_resources(mdev);
return rc;
}
static int __devexit mace_remove(struct macio_dev *mdev)
{
struct net_device *dev = macio_get_drvdata(mdev);
struct mace_data *mp;
BUG_ON(dev == NULL);
macio_set_drvdata(mdev, NULL);
mp = netdev_priv(dev);
unregister_netdev(dev);
free_irq(dev->irq, dev);
free_irq(mp->tx_dma_intr, dev);
free_irq(mp->rx_dma_intr, dev);
iounmap(mp->rx_dma);
iounmap(mp->tx_dma);
iounmap(mp->mace);
free_netdev(dev);
macio_release_resources(mdev);
return 0;
}
static void dbdma_reset(volatile struct dbdma_regs __iomem *dma)
{
int i;
out_le32(&dma->control, (WAKE|FLUSH|PAUSE|RUN) << 16);
/*
* Yes this looks peculiar, but apparently it needs to be this
* way on some machines.
*/
for (i = 200; i > 0; --i)
if (ld_le32(&dma->control) & RUN)
udelay(1);
}
static void mace_reset(struct net_device *dev)
{
struct mace_data *mp = netdev_priv(dev);
volatile struct mace __iomem *mb = mp->mace;
int i;
/* soft-reset the chip */
i = 200;
while (--i) {
out_8(&mb->biucc, SWRST);
if (in_8(&mb->biucc) & SWRST) {
udelay(10);
continue;
}
break;
}
if (!i) {
printk(KERN_ERR "mace: cannot reset chip!\n");
return;
}
out_8(&mb->imr, 0xff); /* disable all intrs for now */
i = in_8(&mb->ir);
out_8(&mb->maccc, 0); /* turn off tx, rx */
out_8(&mb->biucc, XMTSP_64);
out_8(&mb->utr, RTRD);
out_8(&mb->fifocc, RCVFW_32 | XMTFW_16 | XMTFWU | RCVFWU | XMTBRST);
out_8(&mb->xmtfc, AUTO_PAD_XMIT); /* auto-pad short frames */
out_8(&mb->rcvfc, 0);
/* load up the hardware address */
__mace_set_address(dev, dev->dev_addr);
/* clear the multicast filter */
if (mp->chipid == BROKEN_ADDRCHG_REV)
out_8(&mb->iac, LOGADDR);
else {
out_8(&mb->iac, ADDRCHG | LOGADDR);
while ((in_8(&mb->iac) & ADDRCHG) != 0)
;
}
for (i = 0; i < 8; ++i)
out_8(&mb->ladrf, 0);
/* done changing address */
if (mp->chipid != BROKEN_ADDRCHG_REV)
out_8(&mb->iac, 0);
if (mp->port_aaui)
out_8(&mb->plscc, PORTSEL_AUI + ENPLSIO);
else
out_8(&mb->plscc, PORTSEL_GPSI + ENPLSIO);
}
static void __mace_set_address(struct net_device *dev, void *addr)
{
struct mace_data *mp = netdev_priv(dev);
volatile struct mace __iomem *mb = mp->mace;
unsigned char *p = addr;
int i;
/* load up the hardware address */
if (mp->chipid == BROKEN_ADDRCHG_REV)
out_8(&mb->iac, PHYADDR);
else {
out_8(&mb->iac, ADDRCHG | PHYADDR);
while ((in_8(&mb->iac) & ADDRCHG) != 0)
;
}
for (i = 0; i < 6; ++i)
out_8(&mb->padr, dev->dev_addr[i] = p[i]);
if (mp->chipid != BROKEN_ADDRCHG_REV)
out_8(&mb->iac, 0);
}
static int mace_set_address(struct net_device *dev, void *addr)
{
struct mace_data *mp = netdev_priv(dev);
volatile struct mace __iomem *mb = mp->mace;
unsigned long flags;
spin_lock_irqsave(&mp->lock, flags);
__mace_set_address(dev, addr);
/* note: setting ADDRCHG clears ENRCV */
out_8(&mb->maccc, mp->maccc);
spin_unlock_irqrestore(&mp->lock, flags);
return 0;
}
static inline void mace_clean_rings(struct mace_data *mp)
{
int i;
/* free some skb's */
for (i = 0; i < N_RX_RING; ++i) {
if (mp->rx_bufs[i] != NULL) {
dev_kfree_skb(mp->rx_bufs[i]);
mp->rx_bufs[i] = NULL;
}
}
for (i = mp->tx_empty; i != mp->tx_fill; ) {
dev_kfree_skb(mp->tx_bufs[i]);
if (++i >= N_TX_RING)
i = 0;
}
}
static int mace_open(struct net_device *dev)
{
struct mace_data *mp = netdev_priv(dev);
volatile struct mace __iomem *mb = mp->mace;
volatile struct dbdma_regs __iomem *rd = mp->rx_dma;
volatile struct dbdma_regs __iomem *td = mp->tx_dma;
volatile struct dbdma_cmd *cp;
int i;
struct sk_buff *skb;
unsigned char *data;
/* reset the chip */
mace_reset(dev);
/* initialize list of sk_buffs for receiving and set up recv dma */
mace_clean_rings(mp);
memset((char *)mp->rx_cmds, 0, N_RX_RING * sizeof(struct dbdma_cmd));
cp = mp->rx_cmds;
for (i = 0; i < N_RX_RING - 1; ++i) {
skb = dev_alloc_skb(RX_BUFLEN + 2);
if (!skb) {
data = dummy_buf;
} else {
skb_reserve(skb, 2); /* so IP header lands on 4-byte bdry */
data = skb->data;
}
mp->rx_bufs[i] = skb;
st_le16(&cp->req_count, RX_BUFLEN);
st_le16(&cp->command, INPUT_LAST + INTR_ALWAYS);
st_le32(&cp->phy_addr, virt_to_bus(data));
cp->xfer_status = 0;
++cp;
}
mp->rx_bufs[i] = NULL;
st_le16(&cp->command, DBDMA_STOP);
mp->rx_fill = i;
mp->rx_empty = 0;
/* Put a branch back to the beginning of the receive command list */
++cp;
st_le16(&cp->command, DBDMA_NOP + BR_ALWAYS);
st_le32(&cp->cmd_dep, virt_to_bus(mp->rx_cmds));
/* start rx dma */
out_le32(&rd->control, (RUN|PAUSE|FLUSH|WAKE) << 16); /* clear run bit */
out_le32(&rd->cmdptr, virt_to_bus(mp->rx_cmds));
out_le32(&rd->control, (RUN << 16) | RUN);
/* put a branch at the end of the tx command list */
cp = mp->tx_cmds + NCMDS_TX * N_TX_RING;
st_le16(&cp->command, DBDMA_NOP + BR_ALWAYS);
st_le32(&cp->cmd_dep, virt_to_bus(mp->tx_cmds));
/* reset tx dma */
out_le32(&td->control, (RUN|PAUSE|FLUSH|WAKE) << 16);
out_le32(&td->cmdptr, virt_to_bus(mp->tx_cmds));
mp->tx_fill = 0;
mp->tx_empty = 0;
mp->tx_fullup = 0;
mp->tx_active = 0;
mp->tx_bad_runt = 0;
/* turn it on! */
out_8(&mb->maccc, mp->maccc);
/* enable all interrupts except receive interrupts */
out_8(&mb->imr, RCVINT);
return 0;
}
static int mace_close(struct net_device *dev)
{
struct mace_data *mp = netdev_priv(dev);
volatile struct mace __iomem *mb = mp->mace;
volatile struct dbdma_regs __iomem *rd = mp->rx_dma;
volatile struct dbdma_regs __iomem *td = mp->tx_dma;
/* disable rx and tx */
out_8(&mb->maccc, 0);
out_8(&mb->imr, 0xff); /* disable all intrs */
/* disable rx and tx dma */
st_le32(&rd->control, (RUN|PAUSE|FLUSH|WAKE) << 16); /* clear run bit */
st_le32(&td->control, (RUN|PAUSE|FLUSH|WAKE) << 16); /* clear run bit */
mace_clean_rings(mp);
return 0;
}
static inline void mace_set_timeout(struct net_device *dev)
{
struct mace_data *mp = netdev_priv(dev);
if (mp->timeout_active)
del_timer(&mp->tx_timeout);
mp->tx_timeout.expires = jiffies + TX_TIMEOUT;
mp->tx_timeout.function = mace_tx_timeout;
mp->tx_timeout.data = (unsigned long) dev;
add_timer(&mp->tx_timeout);
mp->timeout_active = 1;
}
static int mace_xmit_start(struct sk_buff *skb, struct net_device *dev)
{
struct mace_data *mp = netdev_priv(dev);
volatile struct dbdma_regs __iomem *td = mp->tx_dma;
volatile struct dbdma_cmd *cp, *np;
unsigned long flags;
int fill, next, len;
/* see if there's a free slot in the tx ring */
spin_lock_irqsave(&mp->lock, flags);
fill = mp->tx_fill;
next = fill + 1;
if (next >= N_TX_RING)
next = 0;
if (next == mp->tx_empty) {
netif_stop_queue(dev);
mp->tx_fullup = 1;
spin_unlock_irqrestore(&mp->lock, flags);
return 1; /* can't take it at the moment */
}
spin_unlock_irqrestore(&mp->lock, flags);
/* partially fill in the dma command block */
len = skb->len;
if (len > ETH_FRAME_LEN) {
printk(KERN_DEBUG "mace: xmit frame too long (%d)\n", len);
len = ETH_FRAME_LEN;
}
mp->tx_bufs[fill] = skb;
cp = mp->tx_cmds + NCMDS_TX * fill;
st_le16(&cp->req_count, len);
st_le32(&cp->phy_addr, virt_to_bus(skb->data));
np = mp->tx_cmds + NCMDS_TX * next;
out_le16(&np->command, DBDMA_STOP);
/* poke the tx dma channel */
spin_lock_irqsave(&mp->lock, flags);
mp->tx_fill = next;
if (!mp->tx_bad_runt && mp->tx_active < MAX_TX_ACTIVE) {
out_le16(&cp->xfer_status, 0);
out_le16(&cp->command, OUTPUT_LAST);
out_le32(&td->control, ((RUN|WAKE) << 16) + (RUN|WAKE));
++mp->tx_active;
mace_set_timeout(dev);
}
if (++next >= N_TX_RING)
next = 0;
if (next == mp->tx_empty)
netif_stop_queue(dev);
spin_unlock_irqrestore(&mp->lock, flags);
return 0;
}
static void mace_set_multicast(struct net_device *dev)
{
struct mace_data *mp = netdev_priv(dev);
volatile struct mace __iomem *mb = mp->mace;
int i, j;
u32 crc;
unsigned long flags;
spin_lock_irqsave(&mp->lock, flags);
mp->maccc &= ~PROM;
if (dev->flags & IFF_PROMISC) {
mp->maccc |= PROM;
} else {
unsigned char multicast_filter[8];
struct dev_mc_list *dmi = dev->mc_list;
if (dev->flags & IFF_ALLMULTI) {
for (i = 0; i < 8; i++)
multicast_filter[i] = 0xff;
} else {
for (i = 0; i < 8; i++)
multicast_filter[i] = 0;
for (i = 0; i < dev->mc_count; i++) {
crc = ether_crc_le(6, dmi->dmi_addr);
j = crc >> 26; /* bit number in multicast_filter */
multicast_filter[j >> 3] |= 1 << (j & 7);
dmi = dmi->next;
}
}
#if 0
printk("Multicast filter :");
for (i = 0; i < 8; i++)
printk("%02x ", multicast_filter[i]);
printk("\n");
#endif
if (mp->chipid == BROKEN_ADDRCHG_REV)
out_8(&mb->iac, LOGADDR);
else {
out_8(&mb->iac, ADDRCHG | LOGADDR);
while ((in_8(&mb->iac) & ADDRCHG) != 0)
;
}
for (i = 0; i < 8; ++i)
out_8(&mb->ladrf, multicast_filter[i]);
if (mp->chipid != BROKEN_ADDRCHG_REV)
out_8(&mb->iac, 0);
}
/* reset maccc */
out_8(&mb->maccc, mp->maccc);
spin_unlock_irqrestore(&mp->lock, flags);
}
static void mace_handle_misc_intrs(struct mace_data *mp, int intr, struct net_device *dev)
{
volatile struct mace __iomem *mb = mp->mace;
static int mace_babbles, mace_jabbers;
if (intr & MPCO)
dev->stats.rx_missed_errors += 256;
dev->stats.rx_missed_errors += in_8(&mb->mpc); /* reading clears it */
if (intr & RNTPCO)
dev->stats.rx_length_errors += 256;
dev->stats.rx_length_errors += in_8(&mb->rntpc); /* reading clears it */
if (intr & CERR)
++dev->stats.tx_heartbeat_errors;
if (intr & BABBLE)
if (mace_babbles++ < 4)
printk(KERN_DEBUG "mace: babbling transmitter\n");
if (intr & JABBER)
if (mace_jabbers++ < 4)
printk(KERN_DEBUG "mace: jabbering transceiver\n");
}
IRQ: Maintain regs pointer globally rather than passing to IRQ handlers Maintain a per-CPU global "struct pt_regs *" variable which can be used instead of passing regs around manually through all ~1800 interrupt handlers in the Linux kernel. The regs pointer is used in few places, but it potentially costs both stack space and code to pass it around. On the FRV arch, removing the regs parameter from all the genirq function results in a 20% speed up of the IRQ exit path (ie: from leaving timer_interrupt() to leaving do_IRQ()). Where appropriate, an arch may override the generic storage facility and do something different with the variable. On FRV, for instance, the address is maintained in GR28 at all times inside the kernel as part of general exception handling. Having looked over the code, it appears that the parameter may be handed down through up to twenty or so layers of functions. Consider a USB character device attached to a USB hub, attached to a USB controller that posts its interrupts through a cascaded auxiliary interrupt controller. A character device driver may want to pass regs to the sysrq handler through the input layer which adds another few layers of parameter passing. I've build this code with allyesconfig for x86_64 and i386. I've runtested the main part of the code on FRV and i386, though I can't test most of the drivers. I've also done partial conversion for powerpc and MIPS - these at least compile with minimal configurations. This will affect all archs. Mostly the changes should be relatively easy. Take do_IRQ(), store the regs pointer at the beginning, saving the old one: struct pt_regs *old_regs = set_irq_regs(regs); And put the old one back at the end: set_irq_regs(old_regs); Don't pass regs through to generic_handle_irq() or __do_IRQ(). In timer_interrupt(), this sort of change will be necessary: - update_process_times(user_mode(regs)); - profile_tick(CPU_PROFILING, regs); + update_process_times(user_mode(get_irq_regs())); + profile_tick(CPU_PROFILING); I'd like to move update_process_times()'s use of get_irq_regs() into itself, except that i386, alone of the archs, uses something other than user_mode(). Some notes on the interrupt handling in the drivers: (*) input_dev() is now gone entirely. The regs pointer is no longer stored in the input_dev struct. (*) finish_unlinks() in drivers/usb/host/ohci-q.c needs checking. It does something different depending on whether it's been supplied with a regs pointer or not. (*) Various IRQ handler function pointers have been moved to type irq_handler_t. Signed-Off-By: David Howells <dhowells@redhat.com> (cherry picked from 1b16e7ac850969f38b375e511e3fa2f474a33867 commit)
2006-10-05 21:55:46 +08:00
static irqreturn_t mace_interrupt(int irq, void *dev_id)
{
struct net_device *dev = (struct net_device *) dev_id;
struct mace_data *mp = netdev_priv(dev);
volatile struct mace __iomem *mb = mp->mace;
volatile struct dbdma_regs __iomem *td = mp->tx_dma;
volatile struct dbdma_cmd *cp;
int intr, fs, i, stat, x;
int xcount, dstat;
unsigned long flags;
/* static int mace_last_fs, mace_last_xcount; */
spin_lock_irqsave(&mp->lock, flags);
intr = in_8(&mb->ir); /* read interrupt register */
in_8(&mb->xmtrc); /* get retries */
mace_handle_misc_intrs(mp, intr, dev);
i = mp->tx_empty;
while (in_8(&mb->pr) & XMTSV) {
del_timer(&mp->tx_timeout);
mp->timeout_active = 0;
/*
* Clear any interrupt indication associated with this status
* word. This appears to unlatch any error indication from
* the DMA controller.
*/
intr = in_8(&mb->ir);
if (intr != 0)
mace_handle_misc_intrs(mp, intr, dev);
if (mp->tx_bad_runt) {
fs = in_8(&mb->xmtfs);
mp->tx_bad_runt = 0;
out_8(&mb->xmtfc, AUTO_PAD_XMIT);
continue;
}
dstat = ld_le32(&td->status);
/* stop DMA controller */
out_le32(&td->control, RUN << 16);
/*
* xcount is the number of complete frames which have been
* written to the fifo but for which status has not been read.
*/
xcount = (in_8(&mb->fifofc) >> XMTFC_SH) & XMTFC_MASK;
if (xcount == 0 || (dstat & DEAD)) {
/*
* If a packet was aborted before the DMA controller has
* finished transferring it, it seems that there are 2 bytes
* which are stuck in some buffer somewhere. These will get
* transmitted as soon as we read the frame status (which
* reenables the transmit data transfer request). Turning
* off the DMA controller and/or resetting the MACE doesn't
* help. So we disable auto-padding and FCS transmission
* so the two bytes will only be a runt packet which should
* be ignored by other stations.
*/
out_8(&mb->xmtfc, DXMTFCS);
}
fs = in_8(&mb->xmtfs);
if ((fs & XMTSV) == 0) {
printk(KERN_ERR "mace: xmtfs not valid! (fs=%x xc=%d ds=%x)\n",
fs, xcount, dstat);
mace_reset(dev);
/*
* XXX mace likes to hang the machine after a xmtfs error.
* This is hard to reproduce, reseting *may* help
*/
}
cp = mp->tx_cmds + NCMDS_TX * i;
stat = ld_le16(&cp->xfer_status);
if ((fs & (UFLO|LCOL|LCAR|RTRY)) || (dstat & DEAD) || xcount == 0) {
/*
* Check whether there were in fact 2 bytes written to
* the transmit FIFO.
*/
udelay(1);
x = (in_8(&mb->fifofc) >> XMTFC_SH) & XMTFC_MASK;
if (x != 0) {
/* there were two bytes with an end-of-packet indication */
mp->tx_bad_runt = 1;
mace_set_timeout(dev);
} else {
/*
* Either there weren't the two bytes buffered up, or they
* didn't have an end-of-packet indication.
* We flush the transmit FIFO just in case (by setting the
* XMTFWU bit with the transmitter disabled).
*/
out_8(&mb->maccc, in_8(&mb->maccc) & ~ENXMT);
out_8(&mb->fifocc, in_8(&mb->fifocc) | XMTFWU);
udelay(1);
out_8(&mb->maccc, in_8(&mb->maccc) | ENXMT);
out_8(&mb->xmtfc, AUTO_PAD_XMIT);
}
}
/* dma should have finished */
if (i == mp->tx_fill) {
printk(KERN_DEBUG "mace: tx ring ran out? (fs=%x xc=%d ds=%x)\n",
fs, xcount, dstat);
continue;
}
/* Update stats */
if (fs & (UFLO|LCOL|LCAR|RTRY)) {
++dev->stats.tx_errors;
if (fs & LCAR)
++dev->stats.tx_carrier_errors;
if (fs & (UFLO|LCOL|RTRY))
++dev->stats.tx_aborted_errors;
} else {
dev->stats.tx_bytes += mp->tx_bufs[i]->len;
++dev->stats.tx_packets;
}
dev_kfree_skb_irq(mp->tx_bufs[i]);
--mp->tx_active;
if (++i >= N_TX_RING)
i = 0;
#if 0
mace_last_fs = fs;
mace_last_xcount = xcount;
#endif
}
if (i != mp->tx_empty) {
mp->tx_fullup = 0;
netif_wake_queue(dev);
}
mp->tx_empty = i;
i += mp->tx_active;
if (i >= N_TX_RING)
i -= N_TX_RING;
if (!mp->tx_bad_runt && i != mp->tx_fill && mp->tx_active < MAX_TX_ACTIVE) {
do {
/* set up the next one */
cp = mp->tx_cmds + NCMDS_TX * i;
out_le16(&cp->xfer_status, 0);
out_le16(&cp->command, OUTPUT_LAST);
++mp->tx_active;
if (++i >= N_TX_RING)
i = 0;
} while (i != mp->tx_fill && mp->tx_active < MAX_TX_ACTIVE);
out_le32(&td->control, ((RUN|WAKE) << 16) + (RUN|WAKE));
mace_set_timeout(dev);
}
spin_unlock_irqrestore(&mp->lock, flags);
return IRQ_HANDLED;
}
static void mace_tx_timeout(unsigned long data)
{
struct net_device *dev = (struct net_device *) data;
struct mace_data *mp = netdev_priv(dev);
volatile struct mace __iomem *mb = mp->mace;
volatile struct dbdma_regs __iomem *td = mp->tx_dma;
volatile struct dbdma_regs __iomem *rd = mp->rx_dma;
volatile struct dbdma_cmd *cp;
unsigned long flags;
int i;
spin_lock_irqsave(&mp->lock, flags);
mp->timeout_active = 0;
if (mp->tx_active == 0 && !mp->tx_bad_runt)
goto out;
/* update various counters */
mace_handle_misc_intrs(mp, in_8(&mb->ir), dev);
cp = mp->tx_cmds + NCMDS_TX * mp->tx_empty;
/* turn off both tx and rx and reset the chip */
out_8(&mb->maccc, 0);
printk(KERN_ERR "mace: transmit timeout - resetting\n");
dbdma_reset(td);
mace_reset(dev);
/* restart rx dma */
cp = bus_to_virt(ld_le32(&rd->cmdptr));
dbdma_reset(rd);
out_le16(&cp->xfer_status, 0);
out_le32(&rd->cmdptr, virt_to_bus(cp));
out_le32(&rd->control, (RUN << 16) | RUN);
/* fix up the transmit side */
i = mp->tx_empty;
mp->tx_active = 0;
++dev->stats.tx_errors;
if (mp->tx_bad_runt) {
mp->tx_bad_runt = 0;
} else if (i != mp->tx_fill) {
dev_kfree_skb(mp->tx_bufs[i]);
if (++i >= N_TX_RING)
i = 0;
mp->tx_empty = i;
}
mp->tx_fullup = 0;
netif_wake_queue(dev);
if (i != mp->tx_fill) {
cp = mp->tx_cmds + NCMDS_TX * i;
out_le16(&cp->xfer_status, 0);
out_le16(&cp->command, OUTPUT_LAST);
out_le32(&td->cmdptr, virt_to_bus(cp));
out_le32(&td->control, (RUN << 16) | RUN);
++mp->tx_active;
mace_set_timeout(dev);
}
/* turn it back on */
out_8(&mb->imr, RCVINT);
out_8(&mb->maccc, mp->maccc);
out:
spin_unlock_irqrestore(&mp->lock, flags);
}
IRQ: Maintain regs pointer globally rather than passing to IRQ handlers Maintain a per-CPU global "struct pt_regs *" variable which can be used instead of passing regs around manually through all ~1800 interrupt handlers in the Linux kernel. The regs pointer is used in few places, but it potentially costs both stack space and code to pass it around. On the FRV arch, removing the regs parameter from all the genirq function results in a 20% speed up of the IRQ exit path (ie: from leaving timer_interrupt() to leaving do_IRQ()). Where appropriate, an arch may override the generic storage facility and do something different with the variable. On FRV, for instance, the address is maintained in GR28 at all times inside the kernel as part of general exception handling. Having looked over the code, it appears that the parameter may be handed down through up to twenty or so layers of functions. Consider a USB character device attached to a USB hub, attached to a USB controller that posts its interrupts through a cascaded auxiliary interrupt controller. A character device driver may want to pass regs to the sysrq handler through the input layer which adds another few layers of parameter passing. I've build this code with allyesconfig for x86_64 and i386. I've runtested the main part of the code on FRV and i386, though I can't test most of the drivers. I've also done partial conversion for powerpc and MIPS - these at least compile with minimal configurations. This will affect all archs. Mostly the changes should be relatively easy. Take do_IRQ(), store the regs pointer at the beginning, saving the old one: struct pt_regs *old_regs = set_irq_regs(regs); And put the old one back at the end: set_irq_regs(old_regs); Don't pass regs through to generic_handle_irq() or __do_IRQ(). In timer_interrupt(), this sort of change will be necessary: - update_process_times(user_mode(regs)); - profile_tick(CPU_PROFILING, regs); + update_process_times(user_mode(get_irq_regs())); + profile_tick(CPU_PROFILING); I'd like to move update_process_times()'s use of get_irq_regs() into itself, except that i386, alone of the archs, uses something other than user_mode(). Some notes on the interrupt handling in the drivers: (*) input_dev() is now gone entirely. The regs pointer is no longer stored in the input_dev struct. (*) finish_unlinks() in drivers/usb/host/ohci-q.c needs checking. It does something different depending on whether it's been supplied with a regs pointer or not. (*) Various IRQ handler function pointers have been moved to type irq_handler_t. Signed-Off-By: David Howells <dhowells@redhat.com> (cherry picked from 1b16e7ac850969f38b375e511e3fa2f474a33867 commit)
2006-10-05 21:55:46 +08:00
static irqreturn_t mace_txdma_intr(int irq, void *dev_id)
{
return IRQ_HANDLED;
}
IRQ: Maintain regs pointer globally rather than passing to IRQ handlers Maintain a per-CPU global "struct pt_regs *" variable which can be used instead of passing regs around manually through all ~1800 interrupt handlers in the Linux kernel. The regs pointer is used in few places, but it potentially costs both stack space and code to pass it around. On the FRV arch, removing the regs parameter from all the genirq function results in a 20% speed up of the IRQ exit path (ie: from leaving timer_interrupt() to leaving do_IRQ()). Where appropriate, an arch may override the generic storage facility and do something different with the variable. On FRV, for instance, the address is maintained in GR28 at all times inside the kernel as part of general exception handling. Having looked over the code, it appears that the parameter may be handed down through up to twenty or so layers of functions. Consider a USB character device attached to a USB hub, attached to a USB controller that posts its interrupts through a cascaded auxiliary interrupt controller. A character device driver may want to pass regs to the sysrq handler through the input layer which adds another few layers of parameter passing. I've build this code with allyesconfig for x86_64 and i386. I've runtested the main part of the code on FRV and i386, though I can't test most of the drivers. I've also done partial conversion for powerpc and MIPS - these at least compile with minimal configurations. This will affect all archs. Mostly the changes should be relatively easy. Take do_IRQ(), store the regs pointer at the beginning, saving the old one: struct pt_regs *old_regs = set_irq_regs(regs); And put the old one back at the end: set_irq_regs(old_regs); Don't pass regs through to generic_handle_irq() or __do_IRQ(). In timer_interrupt(), this sort of change will be necessary: - update_process_times(user_mode(regs)); - profile_tick(CPU_PROFILING, regs); + update_process_times(user_mode(get_irq_regs())); + profile_tick(CPU_PROFILING); I'd like to move update_process_times()'s use of get_irq_regs() into itself, except that i386, alone of the archs, uses something other than user_mode(). Some notes on the interrupt handling in the drivers: (*) input_dev() is now gone entirely. The regs pointer is no longer stored in the input_dev struct. (*) finish_unlinks() in drivers/usb/host/ohci-q.c needs checking. It does something different depending on whether it's been supplied with a regs pointer or not. (*) Various IRQ handler function pointers have been moved to type irq_handler_t. Signed-Off-By: David Howells <dhowells@redhat.com> (cherry picked from 1b16e7ac850969f38b375e511e3fa2f474a33867 commit)
2006-10-05 21:55:46 +08:00
static irqreturn_t mace_rxdma_intr(int irq, void *dev_id)
{
struct net_device *dev = (struct net_device *) dev_id;
struct mace_data *mp = netdev_priv(dev);
volatile struct dbdma_regs __iomem *rd = mp->rx_dma;
volatile struct dbdma_cmd *cp, *np;
int i, nb, stat, next;
struct sk_buff *skb;
unsigned frame_status;
static int mace_lost_status;
unsigned char *data;
unsigned long flags;
spin_lock_irqsave(&mp->lock, flags);
for (i = mp->rx_empty; i != mp->rx_fill; ) {
cp = mp->rx_cmds + i;
stat = ld_le16(&cp->xfer_status);
if ((stat & ACTIVE) == 0) {
next = i + 1;
if (next >= N_RX_RING)
next = 0;
np = mp->rx_cmds + next;
if (next != mp->rx_fill
&& (ld_le16(&np->xfer_status) & ACTIVE) != 0) {
printk(KERN_DEBUG "mace: lost a status word\n");
++mace_lost_status;
} else
break;
}
nb = ld_le16(&cp->req_count) - ld_le16(&cp->res_count);
out_le16(&cp->command, DBDMA_STOP);
/* got a packet, have a look at it */
skb = mp->rx_bufs[i];
if (!skb) {
++dev->stats.rx_dropped;
} else if (nb > 8) {
data = skb->data;
frame_status = (data[nb-3] << 8) + data[nb-4];
if (frame_status & (RS_OFLO|RS_CLSN|RS_FRAMERR|RS_FCSERR)) {
++dev->stats.rx_errors;
if (frame_status & RS_OFLO)
++dev->stats.rx_over_errors;
if (frame_status & RS_FRAMERR)
++dev->stats.rx_frame_errors;
if (frame_status & RS_FCSERR)
++dev->stats.rx_crc_errors;
} else {
/* Mace feature AUTO_STRIP_RCV is on by default, dropping the
* FCS on frames with 802.3 headers. This means that Ethernet
* frames have 8 extra octets at the end, while 802.3 frames
* have only 4. We need to correctly account for this. */
if (*(unsigned short *)(data+12) < 1536) /* 802.3 header */
nb -= 4;
else /* Ethernet header; mace includes FCS */
nb -= 8;
skb_put(skb, nb);
skb->protocol = eth_type_trans(skb, dev);
dev->stats.rx_bytes += skb->len;
netif_rx(skb);
mp->rx_bufs[i] = NULL;
++dev->stats.rx_packets;
}
} else {
++dev->stats.rx_errors;
++dev->stats.rx_length_errors;
}
/* advance to next */
if (++i >= N_RX_RING)
i = 0;
}
mp->rx_empty = i;
i = mp->rx_fill;
for (;;) {
next = i + 1;
if (next >= N_RX_RING)
next = 0;
if (next == mp->rx_empty)
break;
cp = mp->rx_cmds + i;
skb = mp->rx_bufs[i];
if (!skb) {
skb = dev_alloc_skb(RX_BUFLEN + 2);
if (skb) {
skb_reserve(skb, 2);
mp->rx_bufs[i] = skb;
}
}
st_le16(&cp->req_count, RX_BUFLEN);
data = skb? skb->data: dummy_buf;
st_le32(&cp->phy_addr, virt_to_bus(data));
out_le16(&cp->xfer_status, 0);
out_le16(&cp->command, INPUT_LAST + INTR_ALWAYS);
#if 0
if ((ld_le32(&rd->status) & ACTIVE) != 0) {
out_le32(&rd->control, (PAUSE << 16) | PAUSE);
while ((in_le32(&rd->status) & ACTIVE) != 0)
;
}
#endif
i = next;
}
if (i != mp->rx_fill) {
out_le32(&rd->control, ((RUN|WAKE) << 16) | (RUN|WAKE));
mp->rx_fill = i;
}
spin_unlock_irqrestore(&mp->lock, flags);
return IRQ_HANDLED;
}
static struct of_device_id mace_match[] =
{
{
.name = "mace",
},
{},
};
MODULE_DEVICE_TABLE (of, mace_match);
static struct macio_driver mace_driver =
{
.name = "mace",
.match_table = mace_match,
.probe = mace_probe,
.remove = mace_remove,
};
static int __init mace_init(void)
{
return macio_register_driver(&mace_driver);
}
static void __exit mace_cleanup(void)
{
macio_unregister_driver(&mace_driver);
kfree(dummy_buf);
dummy_buf = NULL;
}
MODULE_AUTHOR("Paul Mackerras");
MODULE_DESCRIPTION("PowerMac MACE driver.");
module_param(port_aaui, int, 0);
MODULE_PARM_DESC(port_aaui, "MACE uses AAUI port (0-1)");
MODULE_LICENSE("GPL");
module_init(mace_init);
module_exit(mace_cleanup);