Re: [PATCH 1/2] workqueue: Catch more locking problems with flush_work()

[Tejun Heo <tj@kernel.org>]

Hello,

On Thu, Apr 19, 2012 at 11:10:41AM -0700, Stephen Boyd wrote:
> On 04/19/12 08:28, Tejun Heo wrote:
> > On Wed, Apr 18, 2012 at 08:25:57PM -0700, Stephen Boyd wrote:
> >> @@ -2513,8 +2513,11 @@ bool flush_work(struct work_struct *work)
> >>  		wait_for_completion(&barr.done);
> >>  		destroy_work_on_stack(&barr.work);
> >>  		return true;
> >> -	} else
> >> +	} else {
> >> +		lock_map_acquire(&work->lockdep_map);
> >> +		lock_map_release(&work->lockdep_map);
> >>  		return false;
> > We don't have this annotation when start_flush_work() succeeds either,
> > right?  IOW, would lockdep trigger when an actual deadlock happens?
> 
> I believe it does although I haven't tested it.

How does it do that?  While wq->lockdep_map would be able to detect
some of the chaining, the read acquire paths probably would miss some
other.  In general, wq->lockdep_map is used to express dependencies
regarding workqueue flushing (and the self flushing) and it would
probably be better to express work item dependencies explicitly using
work->lockdep_map even if it becomes redundant through wq->lockdep_map
sometimes.

> > If not, why not add the acquire/release() before flush_work() does
> > anything?
> 
> I was worried about causing false positive lockdep warnings in the case
> that start_flush_work() succeeds and returns true. In that case, lockdep
> is told about the cwq lockdep map:
> 
> static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
>                              bool wait_executing)
> {
> 
>         .....
> 
>         if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
>                 lock_map_acquire(&cwq->wq->lockdep_map);
>         else
>                 lock_map_acquire_read(&cwq->wq->lockdep_map);
> 
> 
> and so if we acquired the work->lockdep_map before the
> cwq->wq->lockdep_map we would get a warning about ABBA between these two
> lockdep maps. At least that is what I'm lead to believe when I look at
> what process_one_work() is doing. Please correct me if I'm wrong.

All that's necessary is acquiring and releasing work->lockdep_map.
There's no need to nest start_flush_work() inside it.  Without
nesting, there's nothing to worry about ABBA lockdeps.

Thanks.

-- 
tejun