On Wed, Aug 30, 2017 at 04:41:17PM +0900, Byungchul Park wrote: > On Wed, Aug 30, 2017 at 11:09:53AM +0900, Byungchul Park wrote:
> > > Signed-off-by: Peter Zijlstra (Intel) <pet...@infradead.org> > > > --- > > > diff --git a/kernel/workqueue.c b/kernel/workqueue.c > > > index c0331891dec1..ab3c0dc8c7ed 100644 > > > --- a/kernel/workqueue.c > > > +++ b/kernel/workqueue.c > > > @@ -2107,14 +2107,14 @@ __acquires(&pool->lock) > > > * Which would create W1->C->W1 dependencies, even though there is no > > > * actual deadlock possible. There are two solutions, using a > > > * read-recursive acquire on the work(queue) 'locks', but this will then > > > - * hit the lockdep limitation on recursive locks, or simly discard > > > + * hit the lockdep limitation on recursive locks, or simply discard > > > * these locks. > > > * > > > * AFAICT there is no possible deadlock scenario between the > > > * flush_work() and complete() primitives (except for single-threaded > > > * workqueues), so hiding them isn't a problem. > > > */ > > > - crossrelease_hist_start(XHLOCK_PROC, true); > > > + lockdep_invariant_state(true); > > > > This is what I am always curious about. It would be ok if you agree with > > removing this work-around after fixing acquire things in wq. But, you > > keep to say this is essencial. > > > > You should focus on what dependencies actually are, than saparating > > contexts unnecessarily. Of course, we have to do it for each work, _BUT_ > > not between outside of work and each work since there might be > > dependencies between them certainly. > > You have never answered it. I'm curious about your answer. If you can't, > I think you have to revert all your patches. All yours are wrong. Because I don't understand what you're on about. And my patches actually work.
