2000-11-14 01:00:22

by Jes Sorensen

[permalink] [raw]
Subject: [patch] acenic driver update

Hi Linus

Here is a patch to bring the acenic driver update with the latest
fixes. It fixes the following issues:

- Driver wasn't being initialized when compiled statically into the
kernel.
- set_macaddr() set the address incorrectly on little endian machines.
- netif_stop_queue() was called in post softnet mode when entering
start_xmit() which was unnecessary. Now it is only set when the
queue is full.
- SET_MODULE_OWNER() from Jeff Garzik.
- ETHTOOL support done in a way so it is backwards compatible.

The patch has been up for testing on the AceNIC list for almost a week
and so far nobody has reported problems with it so I consider it to be
stable.

Jes

--- /home/jes/test/linux-2.4.0-test11-pre4/drivers/net/acenic.c Mon Nov 13 17:42:12 2000
+++ drivers/net/acenic.c Mon Nov 13 19:26:23 2000
@@ -2,7 +2,7 @@
* acenic.c: Linux driver for the Alteon AceNIC Gigabit Ethernet card
* and other Tigon based cards.
*
- * Copyright 1998-2000 by Jes Sorensen, <[email protected]>.
+ * Copyright 1998-2000 by Jes Sorensen, <[email protected]>.
*
* Thanks to Alteon and 3Com for providing hardware and documentation
* enabling me to write this driver.
@@ -39,6 +39,8 @@
* where the driver would disable
* bus master mode if it had to disable
* write and invalidate.
+ * Stephen Hack <[email protected]>: Fixed ace_set_mac_addr for little
+ * endian systems.
*/

#include <linux/config.h>
@@ -55,10 +57,12 @@
#include <linux/init.h>
#include <linux/delay.h>
#include <linux/mm.h>
+#include <linux/sockios.h>

-#undef INDEX_DEBUG
-
+#ifdef SIOCETHTOOL
#include <linux/ethtool.h>
+#endif
+
#include <net/sock.h>
#include <net/ip.h>

@@ -69,6 +73,9 @@
#include <asm/uaccess.h>


+#undef INDEX_DEBUG
+
+
#ifdef CONFIG_ACENIC_OMIT_TIGON_I
#define ACE_IS_TIGON_I(ap) 0
#else
@@ -119,6 +126,15 @@
#define SMP_CACHE_BYTES L1_CACHE_BYTES
#endif

+#ifndef SET_MODULE_OWNER
+#define SET_MODULE_OWNER(dev) {do{} while(0);}
+#define ACE_MOD_INC_USE_COUNT MOD_INC_USE_COUNT
+#define ACE_MOD_DEC_USE_COUNT MOD_DEC_USE_COUNT
+#else
+#define ACE_MOD_INC_USE_COUNT {do{} while(0);}
+#define ACE_MOD_DEC_USE_COUNT {do{} while(0);}
+#endif
+

#if (LINUX_VERSION_CODE < 0x02030d)
#define pci_resource_start(dev, bar) dev->base_address[bar]
@@ -130,10 +146,6 @@
#define net_device device
#endif

-#if (LINUX_VERSION_CODE >= 0x02031b)
-#define NEW_NETINIT
-#endif
-
#if (LINUX_VERSION_CODE < 0x02032a)
typedef u32 dma_addr_t;

@@ -154,10 +166,20 @@
#if (LINUX_VERSION_CODE < 0x02032b)
/*
* SoftNet
+ *
+ * For pre-softnet kernels we need to tell the upper layer not to
+ * re-enter start_xmit() while we are in there. However softnet
+ * guarantees not to enter while we are in there so there is no need
+ * to do the netif_stop_queue() dance unless the transmit queue really
+ * gets stuck. This should also improve performance according to tests
+ * done by Aman Singla.
*/
-#define dev_kfree_skb_irq(a) dev_kfree_skb(a)
-#define netif_wake_queue(dev) clear_bit(0, &dev->tbusy)
-#define netif_stop_queue(dev) set_bit(0, &dev->tbusy)
+#define dev_kfree_skb_irq(a) dev_kfree_skb(a)
+#define netif_wake_queue(dev) clear_bit(0, &dev->tbusy)
+#define netif_stop_queue(dev) set_bit(0, &dev->tbusy)
+#define late_stop_netif_stop_queue(dev) {do{} while(0);}
+#define early_stop_netif_stop_queue(dev) test_and_set_bit(0,&dev->tbusy)
+#define early_stop_netif_wake_queue(dev) netif_wake_queue(dev)

static inline void netif_start_queue(struct net_device *dev)
{
@@ -166,16 +188,22 @@
dev->start = 1;
}

-#define ace_mark_net_bh(foo) mark_bh(foo)
-#define netif_queue_stopped(dev) dev->tbusy
-#define netif_running(dev) dev->start
-#define ace_if_down(dev) {do{dev->start = 0;}while (0);}
+#define ace_mark_net_bh(foo) mark_bh(foo)
+#define netif_queue_stopped(dev) dev->tbusy
+#define netif_running(dev) dev->start
+#define ace_if_down(dev) {do{dev->start = 0;}while (0);}
#else
-#define NET_BH 0
-#define ace_mark_net_bh(foo) {do{} while(0);}
-#define ace_if_down(dev) {do{} while(0);}
+#define NET_BH 0
+#define late_stop_netif_stop_queue(dev) netif_stop_queue(dev)
+#define early_stop_netif_stop_queue(dev) 0
+#define early_stop_netif_wake_queue(dev) {do{} while(0);}
+#define ace_mark_net_bh(foo) {do{} while(0);}
+#define ace_if_down(dev) {do{} while(0);}
#endif

+#if (LINUX_VERSION_CODE >= 0x02031b)
+#define NEW_NETINIT
+#endif

#define ACE_MAX_MOD_PARMS 8
#define BOARD_IDX_STATIC 0
@@ -400,10 +428,10 @@
static int dis_pci_mem_inval[ACE_MAX_MOD_PARMS] = {1, 1, 1, 1, 1, 1, 1, 1};

static char version[] __initdata =
- "acenic.c: v0.47 09/18/2000 Jes Sorensen, [email protected]\n"
+ "acenic.c: v0.48 11/14/2000 Jes Sorensen, [email protected]\n"
" http://home.cern.ch/~jes/gige/acenic.html\n";

-static struct net_device *root_dev;
+static struct net_device *root_dev = NULL;

static int probed __initdata = 0;

@@ -460,6 +488,8 @@
break;
}

+ SET_MODULE_OWNER(dev);
+
if (!dev->priv)
dev->priv = kmalloc(sizeof(*ap), GFP_KERNEL);
if (!dev->priv) {
@@ -615,7 +645,7 @@


#ifdef MODULE
-MODULE_AUTHOR("Jes Sorensen <[email protected]>");
+MODULE_AUTHOR("Jes Sorensen <[email protected]>");
MODULE_DESCRIPTION("AceNIC/3C985/GA620 Gigabit Ethernet driver");
MODULE_PARM(link, "1-" __MODULE_STRING(8) "i");
MODULE_PARM(trace, "1-" __MODULE_STRING(8) "i");
@@ -634,8 +664,8 @@
short i;

while (root_dev) {
- next = ((struct ace_private *)root_dev->priv)->next;
ap = root_dev->priv;
+ next = ap->next;

regs = ap->regs;

@@ -730,8 +760,8 @@
}


-#ifdef MODULE
#if (LINUX_VERSION_CODE < 0x02032a)
+#ifdef MODULE
int init_module(void)
{
return ace_module_init();
@@ -742,11 +772,11 @@
{
ace_module_cleanup();
}
+#endif
#else
module_init(ace_module_init);
module_exit(ace_module_cleanup);
#endif
-#endif


static void ace_free_descriptors(struct net_device *dev)
@@ -2211,7 +2241,7 @@

netif_start_queue(dev);

- MOD_INC_USE_COUNT;
+ ACE_MOD_INC_USE_COUNT;

/*
* Setup the timer
@@ -2294,7 +2324,7 @@

restore_flags(flags);

- MOD_DEC_USE_COUNT;
+ ACE_MOD_DEC_USE_COUNT;
return 0;
}

@@ -2307,14 +2337,10 @@
u32 idx, flagsize;

/*
- * ARGH, there is just no pretty way to do this
+ * This only happens with pre-softnet, ie. 2.2.x kernels.
*/
-#if (LINUX_VERSION_CODE < 0x02032b)
- if (test_and_set_bit(0, &dev->tbusy))
+ if (early_stop_netif_stop_queue(dev))
return 1;
-#else
- netif_stop_queue(dev);
-#endif

idx = ap->tx_prd;

@@ -2358,7 +2384,8 @@
*/
mod_timer(&ap->timer, jiffies + (3 * HZ));

- /* The following check will fix a race between the interrupt
+ /*
+ * The following check will fix a race between the interrupt
* handler increasing the tx_ret_csm and testing for tx_full
* and this tx routine's testing the tx_ret_csm and setting
* the tx_full; note that this fix makes assumptions on the
@@ -2369,13 +2396,17 @@
if (((idx + 2) % TX_RING_ENTRIES != ap->tx_ret_csm)
&& xchg(&ap->tx_full, 0)) {
del_timer(&ap->timer);
+ /*
+ * We may not need this one in the post softnet era
+ * in this case this can be changed to a
+ * early_stop_netif_wake_queue(dev);
+ */
netif_wake_queue(dev);
+ } else {
+ late_stop_netif_stop_queue(dev);
}
} else {
- /*
- * No need for it to be atomic - seems it needs to be
- */
- netif_wake_queue(dev);
+ early_stop_netif_wake_queue(dev);
}

dev->trans_start = jiffies;
@@ -2424,6 +2455,7 @@

static int ace_ioctl(struct net_device *dev, struct ifreq *ifr, int cmd)
{
+#ifdef SIOCETHTOOL
struct ace_private *ap = dev->priv;
struct ace_regs *regs = ap->regs;
struct ethtool_cmd ecmd;
@@ -2483,7 +2515,11 @@
ecmd.autoneg = AUTONEG_DISABLE;

#if 0
+ /*
+ * Current struct ethtool_cmd is insufficient
+ */
ecmd.trace = readl(&regs->TuneTrace);
+
ecmd.txcoal = readl(&regs->TuneTxCoalTicks);
ecmd.rxcoal = readl(&regs->TuneRxCoalTicks);
#endif
@@ -2551,6 +2587,7 @@
}
return 0;
}
+#endif

return -EOPNOTSUPP;
}
@@ -2563,7 +2600,7 @@
{
struct sockaddr *addr=p;
struct ace_regs *regs;
- u16 *da;
+ u8 *da;
struct cmd cmd;

if(netif_running(dev))
@@ -2571,11 +2608,11 @@

memcpy(dev->dev_addr, addr->sa_data,dev->addr_len);

- da = (u16 *)dev->dev_addr;
+ da = (u8 *)dev->dev_addr;

regs = ((struct ace_private *)dev->priv)->regs;
- writel(da[0], &regs->MacAddrHi);
- writel((da[1] << 16) | da[2], &regs->MacAddrLo);
+ writel(da[0] << 8 | da[1], &regs->MacAddrHi);
+ writel((da[2] << 24) | (da[3] << 16) | (da[4] << 8) | da[5] , &regs->MacAddrLo);

cmd.evt = C_SET_MAC_ADDR;
cmd.code = 0;
@@ -3000,6 +3037,6 @@

/*
* Local variables:
- * compile-command: "gcc -D__KERNEL__ -DMODULE -I../../include -Wall -Wstrict-prototypes -O2 -fomit-frame-pointer -pipe -fno-strength-reduce -DMODVERSIONS -include ../../include/linux/modversions.h -c -o acenic.o acenic.c"
+ * compile-command: "gcc -D__SMP__ -D__KERNEL__ -DMODULE -I../../include -Wall -Wstrict-prototypes -O2 -fomit-frame-pointer -pipe -fno-strength-reduce -DMODVERSIONS -include ../../include/linux/modversions.h -c -o acenic.o acenic.c"
* End:
*/
--- /home/jes/test/linux-2.4.0-test11-pre4/drivers/net/acenic_firmware.h Mon Sep 18 18:16:48 2000
+++ drivers/net/acenic_firmware.h Mon Nov 13 17:22:40 2000
@@ -17,10 +17,11 @@
#define tigonFwSbssLen 0x38
#define tigonFwBssAddr 0x00015dd0
#define tigonFwBssLen 0x2080
-u32 tigonFwText[];
-u32 tigonFwData[];
-u32 tigonFwRodata[];
#ifndef CONFIG_ACENIC_OMIT_TIGON_I
+#define tigonFwText 0
+#define tigonFwData 0
+#define tigonFwRodata 0
+#else
/* Generated by genfw.c */
u32 tigonFwText[(MAX_TEXT_LEN/4) + 1] __initdata = {
0x10000003,


2000-11-14 01:42:45

by Jeff Garzik

[permalink] [raw]
Subject: Re: [patch] acenic driver update

You would think it would be easier to maintain a separate 2.2.x and
2.4.x version of acenic.c at this point :)


Jes Sorensen wrote:
> @@ -2307,14 +2337,10 @@
> u32 idx, flagsize;
>
> /*
> - * ARGH, there is just no pretty way to do this
> + * This only happens with pre-softnet, ie. 2.2.x kernels.
> */
> -#if (LINUX_VERSION_CODE < 0x02032b)
> - if (test_and_set_bit(0, &dev->tbusy))
> + if (early_stop_netif_stop_queue(dev))
> return 1;
> -#else
> - netif_stop_queue(dev);
> -#endif
>
> idx = ap->tx_prd;
>
> @@ -2358,7 +2384,8 @@
> */
> mod_timer(&ap->timer, jiffies + (3 * HZ));
>
> - /* The following check will fix a race between the interrupt
> + /*
> + * The following check will fix a race between the interrupt
> * handler increasing the tx_ret_csm and testing for tx_full
> * and this tx routine's testing the tx_ret_csm and setting
> * the tx_full; note that this fix makes assumptions on the
> @@ -2369,13 +2396,17 @@
> if (((idx + 2) % TX_RING_ENTRIES != ap->tx_ret_csm)
> && xchg(&ap->tx_full, 0)) {
> del_timer(&ap->timer);
> + /*
> + * We may not need this one in the post softnet era
> + * in this case this can be changed to a
> + * early_stop_netif_wake_queue(dev);
> + */
> netif_wake_queue(dev);

nope, you don't need that. if you are start_xmit, and you didn't stop
the queue yourself, there's no need to start it, and definitely no need
to wake...

you can probably remove 'tx_full' as well, as it is usually redundant to
tbusy/netif_{start,stop}_queue.

Jeff


--
Jeff Garzik |
Building 1024 | The chief enemy of creativity is "good" sense
MandrakeSoft | -- Picasso

2000-11-14 18:41:31

by Alexey Kuznetsov

[permalink] [raw]
Subject: Re: [patch] acenic driver update

Hello!

> - netif_stop_queue() was called in post softnet mode when entering
> start_xmit() which was unnecessary. Now it is only set when the
> queue is full.

Jes, it is not necessary since linux-2.0 or so...

All the difference with softnet is that you need not _test_ tbusy
on entry. That's all, nothing more.

So, you do:

#if (LINUX_VERSION_CODE < 0x02032b)
if (test_bit(&tbusy, 0))
return 1;
#endif

on entry to start_xmit.

All the rest of code is invariant for 2.0, 2.2, 2.4, provided
netif_* macros are defined for earlier kernels.

Alexey

2000-11-15 05:02:00

by Jes Sorensen

[permalink] [raw]
Subject: Re: [patch] acenic driver update

>>>>> "Val" == Val Henson <[email protected]> writes:

Val> Jes, I just downloaded the 0.48 acenic driver and it still has a
Val> reproducible null dereference bug. Anyone can oops their machine
Val> by doing:

Bugger I think I lost your patch in the noise. Sorry about that, it'll
be in the next version.

Val> ifconfig <gige> mtu 9000 ping -f -s 60000 <remote gige host>
Val> ifconfig <gige> mtu 1500 ping -f -s 60000 <remote gige host>

Val> I don't have a fix for this.

Hmmm could be a firmware issue, I'll need to look at it. It is however
a kind of bug that only root can cause deliberately. Doing ifconfig
mtu foo ; ifconfig mtu bar is a little far from normal operation ;-)

Thanks
Jes

2000-11-15 11:12:32

by Pekka Pietikainen

[permalink] [raw]
Subject: Re: [patch] acenic driver update

On Wed, Nov 15, 2000 at 05:31:27AM +0100, Jes Sorensen wrote:
> >>>>> "Val" == Val Henson <[email protected]> writes:
> Val> Jes, I just downloaded the 0.48 acenic driver and it still has a
> Val> reproducible null dereference bug. Anyone can oops their machine
> Val> by doing:
>
> Bugger I think I lost your patch in the noise. Sorry about that, it'll
> be in the next version.
>
> Val> ifconfig <gige> mtu 9000 ping -f -s 60000 <remote gige host>
> Val> ifconfig <gige> mtu 1500 ping -f -s 60000 <remote gige host>
>
> Val> I don't have a fix for this.
>
> Hmmm could be a firmware issue, I'll need to look at it. It is however
> a kind of bug that only root can cause deliberately. Doing ifconfig
> mtu foo ; ifconfig mtu bar is a little far from normal operation ;-)
It seems like it's caused by the driver trying to
do things while it's still setting up the rings.

static void ace_rx_int(struct net_device *dev, u32 rxretprd, u32 rxretcsm)
{
...
rip = &ap->skb->rx_jumbo_skbuff[skbidx];
...
skb = rip->skb;
skb_put(skb, retdesc->size); /* crash here */
...
}

while the driver might be doing this at the same time:

for (i = 0; i < RX_JUMBO_RING_ENTRIES; i++) {
if (ap->skb->rx_jumbo_skbuff[i].skb) {
ap->rx_jumbo_ring[i].size = 0;
set_aceaddr(&ap->rx_jumbo_ring[i].addr,
dev_kfree_skb(ap->skb->rx_jumbo_skbuff[i].skb);
ap->skb->rx_jumbo_skbuff[i].skb = NULL;
}
}
--
Pekka Pietikainen