On Mon, 28 Sep 2009 20:41:41 +0200
Eric Dumazet <eric.dumazet@gmail.com> wrote:

> I did a bisection and found commit def0a9b2573e00ab0b486cb5382625203ab4c4a6
> was the origin of the problem on my x86_32 machine.
> 
> def0a9b2573e00ab0b486cb5382625203ab4c4a6 is first bad commit
> commit def0a9b2573e00ab0b486cb5382625203ab4c4a6
> Author: Peter Zijlstra <a.p.zijlstra@chello.nl>
> Date:   Fri Sep 18 20:14:01 2009 +0200
> 
>     sched_clock: Make it NMI safe
> 
>     Arjan complained about the suckyness of TSC on modern machines, and
>     asked if we could do something about that for PERF_SAMPLE_TIME.
> 
>     Make cpu_clock() NMI safe by removing the spinlock and using
>     cmpxchg. This also makes it smaller and more robust.
> 
>     Affects architectures that use HAVE_UNSTABLE_SCHED_CLOCK, i.e. IA64
>     and x86.
> 
>     Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
>     LKML-Reference: <new-submission>
>     Signed-off-by: Ingo Molnar <mingo@elte.hu>

Confirmed. The bisect run on my machine gave me the same bad commit.
The new logic in sched_clock_remove seems racy: the old code got the
locks for the sched_clock_data of the local and the remove cpu before
it changed any value. The new code tries to get to the same result with
a single cmpxchg. Bad things happen if two cpus try to update the clock
values crosswise, no?

-- 
blue skies,
   Martin.

"Reality continues to ruin my life." - Calvin.