On Tue, Aug 22, 2017 at 02:14:38PM +0900, Byungchul Park wrote:
> On Mon, Aug 21, 2017 at 05:46:00PM +0200, Peter Zijlstra wrote:
> > Now given the above observance rule and the fact that the below report
> > is from the complete, the thing that happened appears to be:
> >
> >
> > lockdep_map_acquire(&work->lockdep_map)
> > down_write(&A)
> >
> > down_write(&A)
> > wait_for_completion(&C)
> >
> > lockdep_map_acquire(&work_lockdep_map);
> > complete(&C)
> >
> > Which lockdep then puked over because both sides saw the same work
> > class.
> >
> > Byungchul; should we not exclude the work class itself, it appears to me
> > the workqueue code is explicitly parallel, or am I confused again?
>
> Do you mean the lockdep_map_acquire(&work->lockdep_map) used manuallly?
>
> That was introduced by Johannes:
>
> commit 4e6045f134784f4b158b3c0f7a282b04bd816887
> "workqueue: debug flushing deadlocks with lockdep"
>
> I am not sure but, for that purpose, IMHO, we can use a
> lockdep_map_acquire_read() instead, in process_one_work(), can't we?
That wouldn't work. That annotation is to help find deadlocks like:
mutex_lock(&A)
<work>
mutex_lock(&A)
flush_work(&work)
The 'fake' lock connects the lock chain inside the work to the
lock-chain at flush_work() and thus detects these. That's perfectly
fine.
It just seems to confuse the completions stuff... Let me go read Dave's
email and see if I can come up with something.
