Re: [PATCH] memcg: refill_stock for kmem uncharging too

[Shakeel Butt]

On Fri, Apr 19, 2019 at 1:07 PM Roman Gushchin <guro@xxxxxx> wrote:
>
> On Thu, Apr 18, 2019 at 02:42:24PM -0700, Shakeel Butt wrote:
> > The commit 475d0487a2ad ("mm: memcontrol: use per-cpu stocks for socket
> > memory uncharging") added refill_stock() for skmem uncharging path to
> > optimize workloads having high network traffic. Do the same for the kmem
> > uncharging as well. However bypass the refill for offlined memcgs to not
> > cause zombie apocalypse.
> >
> > Signed-off-by: Shakeel Butt <shakeelb@xxxxxxxxxx>
>
> Hello, Shakeel!
>
> > ---
> >  mm/memcontrol.c | 17 ++++++++---------
> >  1 file changed, 8 insertions(+), 9 deletions(-)
> >
> > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > index 2535e54e7989..7b8de091f572 100644
> > --- a/mm/memcontrol.c
> > +++ b/mm/memcontrol.c
> > @@ -178,6 +178,7 @@ struct mem_cgroup_event {
> >
> >  static void mem_cgroup_threshold(struct mem_cgroup *memcg);
> >  static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
> > +static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages);
> >
> >  /* Stuffs for move charges at task migration. */
> >  /*
> > @@ -2097,10 +2098,7 @@ static void drain_stock(struct memcg_stock_pcp *stock)
> >       struct mem_cgroup *old = stock->cached;
> >
> >       if (stock->nr_pages) {
> > -             page_counter_uncharge(&old->memory, stock->nr_pages);
> > -             if (do_memsw_account())
> > -                     page_counter_uncharge(&old->memsw, stock->nr_pages);
> > -             css_put_many(&old->css, stock->nr_pages);
> > +             cancel_charge(old, stock->nr_pages);
> >               stock->nr_pages = 0;
> >       }
> >       stock->cached = NULL;
> > @@ -2133,6 +2131,11 @@ static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
> >       struct memcg_stock_pcp *stock;
> >       unsigned long flags;
> >
> > +     if (unlikely(!mem_cgroup_online(memcg))) {
> > +             cancel_charge(memcg, nr_pages);
> > +             return;
> > +     }
>
> I'm slightly concerned about this part. Do we really need it?
> The number of "zombies" which we can pin is limited by the number of CPUs,
> and it will drop fast if there is any load on the machine.
>
> If we skip offline memcgs, it can slow down charging/uncharging of skmem,
> which might be a problem, if the socket is in active use by an other cgroup.
> Honestly, I'd drop this part.
>

Sure, I will wait for comments from others and then send the v2 without this.

Shakeel