Re: [PATCH] mm: memcontrol: use per-cpu stocks for socket memory uncharging

[Michal Hocko]

On Wed 30-08-17 13:57:29, Roman Gushchin wrote:
> On Wed, Aug 30, 2017 at 02:55:43PM +0200, Michal Hocko wrote:
> > On Wed 30-08-17 13:44:59, Roman Gushchin wrote:
> > > On Wed, Aug 30, 2017 at 02:36:55PM +0200, Michal Hocko wrote:
> > > > On Tue 29-08-17 11:01:50, Roman Gushchin wrote:
> > > > [...]
> > > > > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > > > > index b9cf3cf4a3d0..a69d23082abf 100644
> > > > > --- a/mm/memcontrol.c
> > > > > +++ b/mm/memcontrol.c
> > > > > @@ -1792,6 +1792,9 @@ static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
> > > > >  	}
> > > > >  	stock->nr_pages += nr_pages;
> > > > >  
> > > > > +	if (stock->nr_pages > CHARGE_BATCH)
> > > > > +		drain_stock(stock);
> > > > > +
> > > > >  	local_irq_restore(flags);
> > > > >  }
> > > > 
> > > > Why do we need this? In other words, why cannot we rely on draining we
> > > > already do?
> > > 
> > > The existing draining depends on memory pressure, so to keep
> > > the accounting (which we expose to a user) reasonable accurate
> > > even without memory pressure, we need to limit the size of per-cpu stocks.
> > 
> > Why don't we need this for regular page charges? Or maybe we do but that
> > sounds like a seprate and an unrealted fix to me.
> 
> Because we never refill more than CHARGE_BATCH.

You are right that a single process will not but try_charge is a preemptible
context and so multiple processes might pass consume_stock and then
charge a N*CHARGE_BATCH. But I agree that this is quite unlikely so a
separate patch is probably not worth it.

-- 
Michal Hocko
SUSE Labs