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

From: Peter Zijlstra
Date: Tue Jun 14 2016 - 08:05:05 EST


On Tue, Jun 14, 2016 at 01:52:53PM +0800, Boqun Feng wrote:
> 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@xxxxxxxxxxxxx>
> > > Signed-off-by: Pan Xinhui <xinhui.pan@xxxxxxxxxxxxxxxxxx>
> >
> > 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.

Very good point indeed. And yes, the whole RCpc thing, but also the
tricky wandering store on PPC/ARM64 ACQUIRE makes for lots of 'fun' we
can do without.

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

Indeed, I've been tempted to introduce those before.

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

Ah, if Davidlohr is working on that, brilliant, less work for me ;-)