Re: [patch 13/13] mm: memcontrol: rewrite uncharge API

[Johannes Weiner <hannes@cmpxchg.org>]

On Tue, Jul 15, 2014 at 05:18:18PM +0200, Michal Hocko wrote:
> On Tue 15-07-14 11:09:37, Johannes Weiner wrote:
> > On Tue, Jul 15, 2014 at 04:23:50PM +0200, Michal Hocko wrote:
> > > On Tue 15-07-14 10:25:45, Michal Hocko wrote:
> [...]
> > > > @@ -2760,15 +2752,15 @@ static void commit_charge(struct page *page, struct mem_cgroup *memcg,
> > > >  		spin_unlock_irq(&zone->lru_lock);
> > > >  	}
> > > >  
> > > > -	mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
> > > > -	unlock_page_cgroup(pc);
> > > > -
> > > > +	local_irq_disable();
> > > > +	mem_cgroup_charge_statistics(memcg, page, nr_pages);
> > > >  	/*
> > > >  	 * "charge_statistics" updated event counter. Then, check it.
> > > >  	 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
> > > >  	 * if they exceeds softlimit.
> > > >  	 */
> > > >  	memcg_check_events(memcg, page);
> > > > +	local_irq_enable();
> > > 
> > > preempt_{enable,disbale} should be sufficient for
> > > mem_cgroup_charge_statistics and memcg_check_events no?
> > > The first one is about per-cpu accounting (and that should be atomic
> > > wrt. IRQ on the same CPU) and the later one uses IRQ safe locks down in
> > > mem_cgroup_update_tree.
> > 
> > How could it be atomic wrt. IRQ on the local CPU when IRQs that modify
> > the counters can fire on the local CPU?
> 
> I meant that __this_atomic_add and __this_cpu_inc should be atomic wrt. IRQ.
> We do not care that an IRQ might jump in between two per-cpu operations.
> This is racy from other CPUs anyway.

It's really about a single RMW (+=) being interrupted by an IRQ.
this_cpu_ guarantees IRQ-atomicity, but __this_cpu_ does not.

We could probably migrate this code over to this_cpu and get rid of
the IRQ disabling later, because, as you said, we don't care about
being interrupted between separate counter updates.