Re: workqueue panic in 3.4 kernel

[Tejun Heo <tj@kernel.org>]

Hello, Lei.

On Wed, Mar 06, 2013 at 10:39:15PM +0800, Lei Wen wrote:
> We find a race condition as below:
> CPU0                                                                  CPU1
> timer interrupt happen
> __run_timers
>    __run_timers::spin_lock_irq(&base->lock)
>    __run_timers::spin_unlock_irq(&base->lock)
> 
>    __cancel_work_timer
> 
>    __cancel_work_timer::del_timer
> 
>    __cancel_work_timer::wait_on_work
> 
>    __cancel_work_timer::clear_work_data
>    __run_timers::call_timer_fn(timer, fn, data);
>          delayed_work_timer_fn::get_work_cwq
>    __run_timers::spin_lock_irq(&base->lock)
> 
> It is possible for __cancel_work_timer to be run over cpu1 __BEFORE__
> cpu0 is ready to
> run the timer callback, which is delayed_work_timer_fn in our case.

If del_timer() happens after the timer starts running, del_timer()
would return NULL and try_to_grab_pending() will be called which will
return >=0 iff if successfully steals the PENDING bit (ie. it's the
sole owner of the work item).  If del_timer() happens before the timer
starts running, the timer function would never run.

clear_work_data() happens iff the work item is confirmed to be idle.
At this point, I'm pretty skeptical this is a bug in workqueue itself
and strongly suggest looking at the crashing workqueue user.

Thanks.

-- 
tejun