Re: [PATCH 1/2] mm/slub: optimize alloc/free fastpath by removing preemption on/off

[Davidlohr Bueso <dave@stgolabs.net>]

On Tue, 2015-01-06 at 17:09 +0900, Joonsoo Kim wrote:
> On Mon, Jan 05, 2015 at 07:03:12PM -0800, Davidlohr Bueso wrote:
> > On Mon, 2015-01-05 at 10:36 +0900, Joonsoo Kim wrote:
> > > -	preempt_disable();
> > > -	c = this_cpu_ptr(s->cpu_slab);
> > > +	do {
> > > +		tid = this_cpu_read(s->cpu_slab->tid);
> > > +		c = this_cpu_ptr(s->cpu_slab);
> > > +	} while (IS_ENABLED(CONFIG_PREEMPT) && unlikely(tid != c->tid));
> > > +	barrier();
> > 
> > I don't see the compiler reodering the object/page stores below, since c
> > is updated in the loop anyway. Is this really necessary (same goes for
> > slab_free)? The generated code by gcc 4.8 looks correct without it.
> > Additionally, the implied barriers for preemption control aren't really
> > the same semantics used here (if that is actually the reason why you are
> > using them).
> 
> Hello,
> 
> I'd like to use tid as a pivot so it should be fetched before fetching
> anything on c. Is it impossible even if !CONFIG_PREEMPT without
> barrier()?

You'd need a smp_wmb() in between tid and c in the loop then, which
looks quite unpleasant. All in all disabling preemption isn't really
that expensive, and you should redo your performance number if you go
this way.