Re: [PATCH 1/2] workqueue: skip lockdep wq dependency in cancel_work_sync()

[Johannes Berg <johannes@sipsolutions.net>]

On Tue, 2018-08-21 at 09:08 -0700, Tejun Heo wrote:
> 
> > -static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
> > +static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
> > +			     bool from_cancel)
> >  {
> >  	struct worker *worker = NULL;
> >  	struct worker_pool *pool;
> > @@ -2885,7 +2886,8 @@ static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
> >  	 * workqueues the deadlock happens when the rescuer stalls, blocking
> >  	 * forward progress.
> >  	 */
> > -	if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer) {
> > +	if (!from_cancel &&
> > +	    (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
> >  		lock_map_acquire(&pwq->wq->lockdep_map);
> >  		lock_map_release(&pwq->wq->lockdep_map);
> >  	}
> 
> But this can lead to a deadlock.  I'd much rather err on the side of
> discouraging complex lock dancing around ordered workqueues, no?

What can lead to a deadlock?

Writing out the example again, with the unlock now:

   work1_function() { mutex_lock(&mutex); mutex_unlock(&mutex); }
   work2_function() { /* nothing */ }
 
   other_function() {
     queue_work(ordered_wq, &work1);
     queue_work(ordered_wq, &work2);
     mutex_lock(&mutex);
     cancel_work_sync(&work2);
     mutex_unlock(&mutex);
   }

This shouldn't be able to lead to a deadlock like I had explained:

> In cancel_work_sync(), we can only have one of two cases, even
> with an ordered workqueue:
>  * the work isn't running, just cancelled before it started
>  * the work is running, but then nothing else can be on the
>    workqueue before it

johannes