Re: [RFC PATCH v2] memory-barriers: remove smp_mb__after_unlock_lock()

[Will Deacon <will.deacon@arm.com>]

On Thu, Aug 20, 2015 at 10:45:05AM +0100, Michael Ellerman wrote:
> On Tue, 2015-08-18 at 09:37 +0100, Will Deacon wrote:
> > On Tue, Aug 18, 2015 at 02:50:55AM +0100, Michael Ellerman wrote:
> > > On Mon, 2015-08-17 at 09:57 +0100, Will Deacon wrote:
> > > > On Mon, Aug 17, 2015 at 07:15:01AM +0100, Paul E. McKenney wrote:
> > > > > On Mon, Aug 17, 2015 at 02:06:07PM +1000, Michael Ellerman wrote:
> > > > > > On Wed, 2015-08-12 at 08:43 -0700, Paul E. McKenney wrote:
> > > > > > I thought the end result of this thread was that we didn't *need* to change the
> > > > > > powerpc lock semantics? Or did I read it wrong?
> > > > > > 
> > > > > > ie. the docs now say that RELEASE+ACQUIRE is not a full barrier, which is
> > > > > > consistent with our current implementation.
> > > > > 
> > > > > That change happened about 1.5 years ago, and I thought that the
> > > > > current discussion was about reversing it, based in part on the
> > > > > recent powerpc benchmarks of locking primitives with and without the
> > > > > sync instruction.  But regardless, I clearly cannot remove either the
> > > > > smp_mb__after_unlock_lock() or the powerpc definition of it to be smp_mb()
> > > > > if powerpc unlock/lock is not strengthened.
> > > > 
> > > > Yup. Peter and I would really like to get rid of smp_mb__after_unlock_lock
> > > > entirely, which would mean strengthening the ppc spinlocks. Moving the
> > > > barrier primitive into RCU is a good step to prevent more widespread usage
> > > > of the barrier, but we'd really like to go further if the performance impact
> > > > is deemed acceptable (which is what this thread is about).
> > > 
> > > OK, sorry for completely missing the point, too many balls in the air here.
> > 
> > No problem!
> > 
> > > I'll do some benchmarks and see what we come up with.
> > 
> > Thanks, that sounds great. FWIW, there are multiple ways of implementing
> > the patch (i.e. whether you strengthen lock or unlock). I had a crack at
> > something here, but it's not tested:
> > 
> >   http://marc.info/?l=linux-arch&m=143758379023849&w=2
> 
> Thanks.
> 
> I notice you are not changing PPC_RELEASE_BARRIER, but only the spin unlock
> code. But from my reading of the docs we need to make sure any UNLOCK+LOCK is a
> full barrier, not just spin unlock/lock?
> 
> So don't we need to worry about some of the other locks as well? At least
> rwlock, and mutex fast path?

Hmm, that's a good question. I notice that you don't do any of the SYNC_IO
stuff for any locks other than spinlocks but I don't know whether
smp_mb__after_unlock_lock is similarly limited in scope.

Paul?

Will