On Mon, 11 Aug 2008, David Miller wrote:

From: Peter Zijlstra <redacted>
Date: Mon, 11 Aug 2008 23:09:38 +0200

quoted

On Mon, 2008-08-11 at 14:00 -0700, David Miller wrote:

quoted

From: "John Kacur" <redacted>
Date: Mon, 11 Aug 2008 15:11:46 +0200

quoted

__qdisc_run() calls qdisc_restart() which calls
handle_dev_cpu_collision(skb, dev, q); and then the problem shows up
here:
__get_cpu_var(netdev_rx_stat).cpu_collision++;

The solution is to disable interrupts around the above increment. Here
is an attached patch to do so. (Thank's to Peter Zijlstra for help in
the analysis and dropping the answer in my lap, so if I got it right
it is due to his help, but if I messed it up, then I did that part all
by myself.)

__qdisc_run() always runs in software interrupt context,
so I guess this is some problem with the -rt stuff running
software interrupts in threads?

Hmm, good point - and those threads should be cpu affine on -rt if I'm
not mistaken. Steven, do you happen to remember details?

The key issue is whether those threads run software interrupts
in a compatible environment.  And such a proper environment allows
plain smp_processor_id() without any special preparations.

Yes, we have a softirq thread per CPU. We should have a test in the 
smp_processor_id for rt to not bug if it is called by known "per_cpu" 
threads.

-- Steve