Messages in this thread | | | Subject | Re: [PATCH -v9][RFC] mutex: implement adaptive spinning | From | Peter Zijlstra <> | Date | Tue, 13 Jan 2009 17:21:50 +0100 |
| |
On Tue, 2009-01-13 at 08:16 -0800, Linus Torvalds wrote: > > On Tue, 13 Jan 2009, Peter Zijlstra wrote: > > > > Change mutex contention behaviour such that it will sometimes busy wait on > > acquisition - moving its behaviour closer to that of spinlocks. > > Okey, dokey. Looks reasonable, but I wonder if this part came from v8 and > wasn't intentional: > > > + if (atomic_xchg(&lock->count, -1) == 1) { > > + lock_acquired(&lock->dep_map, ip); > > + mutex_set_owner(lock); > > + preempt_enable(); > > + return 0; > > + } > > Now you're forcing the slow-path on unlock. Maybe it was intentional, > maybe it wasn't. Did you perhaps mean > > if (atomic_cmpxchg(&lock->count, 1, 0) == 1) { > > here? I thought we agreed it was safe, if only because it should be > equivalent to just having done "mutex_trylock()" instead of a "real" lock > sequence.
Yes, that was an 'accident' from -v8, yes we did think the cmpxchg was good, however I did get some spurious lockups on -v7, and I only noticed the thing after I'd done most of the testing, so I decided to let it be for now.
Let me put the cmpxchg back in and see if this is all still good (only 3*2*2 configs to test :-).
| |