On Tue, 2018-08-21 at 10:55 -0700, Tejun Heo wrote:
> > I'm not really sure what you think we might be missing? Am I missing
> > some case where cancel_work_sync() can possibly deadlock? Apart from the
> > issue I addressed in the second patch, obviously.
>
> Ah, that was me being slow. I thought you were skipping the work's
> lockdep_map. I can almost swear we had that before (the part you're
> adding on the second patch). Right, fd1a5b04dfb8 ("workqueue: Remove
> now redundant lock acquisitions wrt. workqueue flushes") removed it
> because it gets propagated through wait_for_completion(). Did we miss
> some cases with that change?
Hmm.
It doesn't seem to be working.
No, ok, actually it probably *does*, but the point is similar to my
issue # 3 before - we don't do any of this unless the work is actually
running, but we really want the lockdep annotation *regardless* of that,
so that we catch the error unconditionally.
So perhaps that commit just needs to be reverted entirely - I'd only
looked at a small subset of it, but the flush_workqueue() case has the
same problem - we only get to the completion when there's something to
flush, not when the workqueue happens to actually be empty. But again,
for lockdep we want to catch *potential* problems, not only *actual*
ones.
The remaining part of the patch I'm not sure I fully understand (removal
of lockdep_init_map_crosslock()), but I suppose if we revert the other
bits we need to revert this as well.
So please drop this patch, but revert Byungchul Park's commit
fd1a5b04dfb8 again, I don't think the lockdep annotations there are
really redundant as I just explained.
johannes
