Re: [RFC][PATCH 3/3] locking,arm64: Introduce cmpwait()

[Boqun Feng <boqun.feng@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 1493 bytes --]

On Sat, Apr 23, 2016 at 04:37:36AM +0200, Peter Zijlstra wrote:
> On Sat, Apr 23, 2016 at 12:08:57AM +0800, Boqun Feng wrote:
> > How about replace smp_rmb() with a smp_acquire_barrier__after_cmpwait()?
> > This barrier is designed to provide an ACQUIRE ordering when combining a
> > cmpwait() .
> 
> That's a horrible name for a barrier :-)
> 

Indeed, and having another special barrier is a pain.

> > And cmpwait() only has minimal ordering guarantee, but if it is actually
> > an ACQUIRE, then the corresponding smp_acquire_barrier__after_cmpwait()
> > is just empty.
> > 
> > We might need this special barrier on ppc, because we can implement it
> > with "isync" given that cmpwait() has control dependency and ctrl+isync
> > is ACQUIRE on ppc.
> > 
> > Thoughts?
> 
> Provide a PPC specific smp_cond_load_acquire() using ISYNC ?

That works, but I should do more investigation on that, I brought this
up because ISYNC may be better than smp_rmb() in two cases, IIUC:

1.	for old systems without LWSYNC, smp_rmb() is actually SYNC,
	which is heavier than ISYNC.

2.	for new systems, isync may be a little faster than lwsync,
	according to Michael Ellerman: http://lkml.kernel.org/g/1437012028.28475.2.camel@ellerman.id.au

but I doubt these two cases will have any observed performance impact,
so the current version of smp_cond_load_acquire() in your patch is fine,
and we can switch to a specific one in the future we want to ;-)

Regards,
Boqun

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 473 bytes --]