Skip to contentSkip to conversation

2.4, b44 transmit timeout

DISCUSSION Jan De Luyck 2 participants 3 messages Sep 8, 2003 – Sep 28, 2003 Networking

lore
Share & tools
lore
#0 lore raw
Quoted 4 lines · Norbert Preining

A-ha! Thanks, I think this should be enough to figure out what the problem is... Looks like the driver doesn't even get the packets pump tries to send, pump is a bit special in the way it bounces the interface up and down when doing its work, that probably triggers a race in b44..

I can also easily recreate this by calling dhcpcd e.g. when the cable isn't in the socket yet. If i attach the cable then I see the interface coming up, going down, and then the NETDEV watchdog message.

Unfortunatly this usually means that dhcpcd goes hanging. ifconfig hangs too if I try to use it, and rebooting must be forced with sysrq and an oops a alt-sysrq-o for poweroff...

Jan

#1 lore raw
Quoted 7 lines · Jan De Luyck
Quoted 4 lines · Norbert Preining

A-ha! Thanks, I think this should be enough to figure out what the problem is... Looks like the driver doesn't even get the packets pump tries to send, pump is a bit special in the way it bounces the interface up and down when doing its work, that probably triggers a race in b44..

Quoted 7 lines

I can also easily recreate this by calling dhcpcd e.g. when the cable isn't in the socket yet. If i attach the cable then I see the interface coming up, going down, and then the NETDEV watchdog message.

Unfortunatly this usually means that dhcpcd goes hanging. ifconfig hangs too if I try to use it, and rebooting must be forced with sysrq and an oops a alt-sysrq-o for poweroff...

Hi

(I was quite far away from civilization for a few days, which is why I've been quiet on this topic for a while :-) )

Culd you try the patch Jeff Garzik posted as a fix for tg3 ifconfig down/up problems, I believe it fixes the same problem since it's a generic thing with NAP

What I believe should happen after applying this patch is that the b44 driver shouldn't hang as a result of doing tricks like the ones mentioned. pump might not work even with it (If I understood correctly, (pump & tg3 don't mix well either due to pump having some wrong assumptions on when it can send packets). dhcpcd definately should work though.

--- SNIP --
Note that people seeing "ifconfig down ... ifconfig up" problems need to apply this patch. (to 2.4.23-pre, too)

Jeff

[-- Attachment #2: patch --]
[-- Type: text/plain, Encoding: 7bit, Size: 0.5K --]

net/core/dev.c Mon Sep 8 18:14:36 2003 +5 -2
diff -Nru a/net/core/dev.c b/net/core/dev.c
@@ -851,7 +851,11 @@
         * engine, but this requires more changes in devices. */
                                                                                
        smp_mb__after_clear_bit(); /* Commit netif_running(). */
-       netif_poll_disable(dev);
+       while (test_bit(__LINK_STATE_RX_SCHED, &dev->state)) {
+               /* No hurry. */
+               current->state = TASK_INTERRUPTIBLE;
+               schedule_timeout(1);
+       }
 
-- SNIP --
 
#2 lore raw
In reply to Jan De Luyck
Quoted 7 lines · Jan De Luyck
Quoted 4 lines · Norbert Preining

A-ha! Thanks, I think this should be enough to figure out what the problem is... Looks like the driver doesn't even get the packets pump tries to send, pump is a bit special in the way it bounces the interface up and down when doing its work, that probably triggers a race in b44..

Quoted 6 lines

I can also easily recreate this by calling dhcpcd e.g. when the cable isn't in the socket yet. If i attach the cable then I see the interface coming up, going down, and then the NETDEV watchdog message.
Unfortunatly this usually means that dhcpcd goes hanging. ifconfig hangs too if I try to use it, and rebooting must be forced with sysrq and an oops a alt-sysrq-o for poweroff...

This could help (tm) :-)

b44.c 2003-09-28 19:37:07.000000000 +0300 renamed from b44.c.orig 2003-09-28 19:36:48.000000000 +0300 +1 -0
@@ -870,6 +870,7 @@
 
        spin_unlock_irq(&bp->lock);
 
+       b44_enable_ints(bp);
        netif_wake_queue(dev);
 }
 

at least I could do some pretty dirty things to the poor chip and it still seems to recover after this patch.

rmmod still has some problems, occasionally I get a "usage count 2" thing (currently running the rawhide 2.4.22-1.2061.nptl, I've seen similar with 2.6.0-test5ish too, I suspect ipv6 might be involved, but sometimes I can rmmod it even with ipv6 loaded so it's a bit random).