On Wed, Jan 14, 2009 at 07:23:12PM +0200, Avi Kivity wrote: > Nick Piggin wrote: > >>(no they're not, Nick's ticket locks still spin on a shared cacheline > >>IIRC -- the MCS locks mentioned could fix this) > >> > > > >It reminds me. I wrote a basic variation of MCS spinlocks a while back. And > >converted dcache lock to use it, which showed large dbench improvements on > >a big machine (of course for different reasons than the dbench improvements > >in this threaed). > > > >http://lkml.org/lkml/2008/8/28/24 > > > >Each "lock" object is sane in size because given set of spin-local queues > >may > >only be used once per lock stack. But any spinlocks within a mutex > >acquisition > >will always be at the bottom of such a stack anyway, by definition. > > > >If you can use any code or concept for your code, that would be great. > > > > Does it make sense to replace 'nest' with a per-cpu counter that's > incremented on each lock? I guest you'd have to search for the value of > nest on unlock, but it would a very short search (typically length 1, 2 > if lock sorting is used to avoid deadlocks). > > I think you'd need to make the lock store the actual node pointer, not > the cpu number, since the values of nest would be different on each cpu. > > That would allow you to replace spinlocks with mcs_locks wholesale. nest could get quite large (and basically is unbounded), though. I think there would definitely be variations (NMCS locks is interesting). But OTOH I think that for _most_ spinlocks, optimising for spinning case is wrong. For some really tricky global spinlocks, yes I think MCS is a good idea. But moreso as a stopgap until hopefully more scalable algorithms are written. -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
