On Thu, Aug 24, 2017 at 05:11:01PM +0900, Byungchul Park wrote: > On Thu, Aug 24, 2017 at 04:37:13PM +0900, Byungchul Park wrote: > > On Thu, Aug 24, 2017 at 03:11:53PM +0900, Byungchul Park wrote: > > > On Wed, Aug 23, 2017 at 07:47:14PM +0200, Peter Zijlstra wrote: > > > > Those are fine and are indeed the flush_work() vs work inversion. > > > > > > > > The two straight forward annotations are: > > > > > > > > flush_work(work) process_one_work(wq, work) > > > > A(work) A(work) > > > > R(work) work->func(work); > > > > R(work) > > > > > > > > Which catches: > > > > > > > > Task-1: work: > > > > > > > > mutex_lock(&A); mutex_lock(&A); > > > > flush_work(work); > > > > > > I'm not sure but, with LOCKDEP_COMPLETE enabled, this issue would > > > automatically be covered w/o additional A(work)/R(work). Right? > > > > > > A(work)/R(work) seem to be used for preventing wait_for_completion() > > > in flush_work() from waiting for the completion forever because of the > > > work using mutex_lock(&A). Am I understanding correctly? > > > > > > If yes, we can use just LOCKDEP_COMPLETE for that purpose. > > > > I'm not familiar with workqueue but, the manual lockdep_map_acquire() in > > workqueue code seems to be introduced to do what LOCKDEP_COMPLETE does > > for wait_for_completion() and complete(). > > > > Wrong? > > As I understand how workqueue code works more, thanks to Peterz, I get > convinced. What they want to detect with acquire(w/wq) is a deadlock > caused by wait_for_completion() mixed with typical locks. > > We have to detect it with _variables_ which it actually waits for, i.e. > completion variable, neither _work_ nor _workqueue_ which we are > currently using.
Please read this more _carefully_. I would be sorry if wrong, _BUT_, it could be a key to solve the issue of workqueue in right way, IIUC.