Re: [PATCH v5 4/4] futex: Avoid taking hb lock if nothing to wakeup

[Davidlohr Bueso <davidlohr@hp.com>]

On Sat, 2014-01-11 at 01:52 -0800, Paul E. McKenney wrote:
[...]
> On Sat, Jan 11, 2014 at 01:49:12AM -0800, Paul E. McKenney wrote:
> > On Thu, Jan 02, 2014 at 07:05:20AM -0800, Davidlohr Bueso wrote:
> > > -	spin_lock(&hb->lock);
> > > +	spin_lock(&hb->lock); /* implies MB (A) */
> > 
> > You need smp_mb__before_spinlock() before the spin_lock() to get a
> > full memory barrier.

Hmmm, the thing we need to guarantee here is that the ticket increment
is visible (which is the same as the smp_mb__after_atomic_inc we used to
have in the original atomic counter approach), so adding a barrier
before the spin_lock call wouldn't serve that. I previously consulted
this with Linus and we can rely on the fact that spin_lock calls already
update the head counter, so spinners are visible even if the lock hasn't
been acquired yet.

> Actually, even that only gets you smp_mb().

I guess you mean smp_wmb() here.

> Unless you are ordering a prior write against a later write here, you
> will need an smp_mb().

Yep.

Thanks for looking into this,
Davidlohr