> 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: NArjan van de Ven <arjan@linux.intel.com>
Signed-off-by: NIngo Molnar <mingo@elte.hu>
Signed-off-by: NJeff Garzik <jeff@garzik.org>