Re: [PATCH] locking/qspinlock: Use atomic_sub_return_release in queued_spin_unlock

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

[-- Attachment #1: Type: text/plain, Size: 1329 bytes --]

On Mon, Jun 13, 2016 at 12:45:23PM -0700, Davidlohr Bueso wrote:
> On Fri, 03 Jun 2016, Pan Xinhui wrote:
> 
> > The existing version uses a heavy barrier while only release semantics
> > is required. So use atomic_sub_return_release instead.
> > 
> > Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org>
> > Signed-off-by: Pan Xinhui <xinhui.pan@linux.vnet.ibm.com>
> 
> I just noticed this change in -tip and, while I know that saving a barrier
> in core spinlock paths is perhaps a worthy exception, I cannot help but
> wonder if this is the begging of the end for smp__{before,after}_atomic().

This is surely a good direction I think, that is using _acquire and
_release primitives to replace those barriers. However, I think we
should do this carefully, because the _acquire and _release primitives
are RCpc because they are on PPC, IOW, a ACQUIRE and RELEASE pair is not
a full barrier nor provides global transivity. I'm worried about there
are some users depending on the full-barrier semantics, which means we
must audit each use carefully before we make the change.

Besides, if we want to do the conversion, we'd better have _acquire and
_release variants for non-value-returning atomic operations.

I remember you were working on those variants. How is that going?

Regards,
Boqun

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