]> err.no Git - linux-2.6/commit
[PATCH] 8139too deadlock fix
authorArjan van de Ven <arjan@infradead.org>
Wed, 5 Jul 2006 13:00:40 +0000 (15:00 +0200)
committerJeff Garzik <jeff@garzik.org>
Wed, 5 Jul 2006 18:29:26 +0000 (14:29 -0400)
commitbce305f4fe779f29d99d414685243f5da0803254
tree67072c37319f3c6b091dec8094d342118fe3fd62
parent20ed7c094dfe33b0e15e8c60f60012b9278631d3
[PATCH] 8139too deadlock fix

> stack backtrace:

>  [<f9099d31>] rtl8139_start_xmit+0xd9/0xff [8139too]
>  [<c11ad5ea>] netpoll_send_skb+0x98/0xea

This seems to be a real deadlock...

So netpoll_send_skb takes the _xmit_lock, which is all nitty gritty
but then rtl8139_start_xmit comes around while that lock is taken, and
does

      spin_unlock_irq(&tp->lock);

which.. enables interrupts and softirqs; this is quite bad because the
xmit lock is taken in softirq context for the watchdog like this:
  [<c1200376>] _spin_lock+0x23/0x32
  [<c11af282>] dev_watchdog+0x14/0xb1
  [<c101dab2>] run_timer_softirq+0xf2/0x14a
  [<c101a691>] __do_softirq+0x55/0xb0
  [<c1004a8d>] do_softirq+0x58/0xbd

Which would deadlock now that the spin_unlock_irq() has enabled
irqs/softirqs while the _xmit_lock is still held.

The patch below turns this into a irqsave/irqrestore pair so that
interrupts don't get enabled unconditionally.

Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Jeff Garzik <jeff@garzik.org>
drivers/net/8139too.c