On Thu, Oct 13, 2016 at 04:28:01PM +0100, Will Deacon wrote: > On Fri, Oct 07, 2016 at 04:52:51PM +0200, Peter Zijlstra wrote:
> > @@ -457,15 +472,20 @@ static bool mutex_optimistic_spin(struct > > * release the lock or go to sleep. > > */ > > owner = __mutex_owner(lock); > > - if (owner && !mutex_spin_on_owner(lock, owner)) > > - break; > > + if (owner) { > > + if (waiter && owner == task) { > > + smp_mb(); /* ACQUIRE */ > > Hmm, is this barrier actually needed? This only happens on the handoff path, > and we take the wait_lock immediately after this succeeds anyway. That > control dependency, coupled with the acquire semantics of the spin_lock, > should be sufficient, no? Yes, I think you're right. But like said in that earlier email, I'd like to keep this for now. Once this code has settled we can reconsider this.