On Sun, Aug 18, 2019 at 3:14 AM Roman Gushchin <guro@xxxxxx> wrote:
>
> On Sat, Aug 17, 2019 at 11:33:57AM +0800, Yafang Shao wrote:
> > On Sat, Aug 17, 2019 at 8:47 AM Roman Gushchin <guro@xxxxxx> wrote:
> > >
> > > Commit 766a4c19d880 ("mm/memcontrol.c: keep local VM counters in sync
> > > with the hierarchical ones") effectively decreased the precision of
> > > per-memcg vmstats_local and per-memcg-per-node lruvec percpu counters.
> > >
> > > That's good for displaying in memory.stat, but brings a serious regression
> > > into the reclaim process.
> > >
> > > One issue I've discovered and debugged is the following:
> > > lruvec_lru_size() can return 0 instead of the actual number of pages
> > > in the lru list, preventing the kernel to reclaim last remaining
> > > pages. Result is yet another dying memory cgroups flooding.
> > > The opposite is also happening: scanning an empty lru list
> > > is the waste of cpu time.
> > >
> > > Also, inactive_list_is_low() can return incorrect values, preventing
> > > the active lru from being scanned and freed. It can fail both because
> > > the size of active and inactive lists are inaccurate, and because
> > > the number of workingset refaults isn't precise. In other words,
> > > the result is pretty random.
> > >
> > > I'm not sure, if using the approximate number of slab pages in
> > > count_shadow_number() is acceptable, but issues described above
> > > are enough to partially revert the patch.
> > >
> > > Let's keep per-memcg vmstat_local batched (they are only used for
> > > displaying stats to the userspace), but keep lruvec stats precise.
> > > This change fixes the dead memcg flooding on my setup.
> > >
> >
> > That will make some misunderstanding if the local counters are not in
> > sync with the hierarchical ones
> > (someone may doubt whether there're something leaked.).
>
> Sure, but the actual leakage is a much more serious issue.
>
> > If we have to do it like this, I think we should better document this behavior.
>
> Lru size calculations can be done using per-zone counters, which is
> actually cheaper, because the number of zones is usually smaller than
> the number of cpus. I'll send a corresponding patch on Monday.
>
Looks like a good idea.
> Maybe other use cases can also be converted?
We'd better keep the behavior the same across counters. I think you
can have a try.
Thanks
Yafang
>
> Thanks!
>
> >
> > > Fixes: 766a4c19d880 ("mm/memcontrol.c: keep local VM counters in sync with the hierarchical ones")
> > > Signed-off-by: Roman Gushchin <guro@xxxxxx>
> > > Cc: Yafang Shao <laoar.shao@xxxxxxxxx>
> > > Cc: Johannes Weiner <hannes@xxxxxxxxxxx>
> > > ---
> > > mm/memcontrol.c | 8 +++-----
> > > 1 file changed, 3 insertions(+), 5 deletions(-)
> > >
> > > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > > index 249187907339..3429340adb56 100644
> > > --- a/mm/memcontrol.c
> > > +++ b/mm/memcontrol.c
> > > @@ -746,15 +746,13 @@ void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
> > > /* Update memcg */
> > > __mod_memcg_state(memcg, idx, val);
> > >
> > > + /* Update lruvec */
> > > + __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
> > > +
> > > x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
> > > if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
> > > struct mem_cgroup_per_node *pi;
> > >
> > > - /*
> > > - * Batch local counters to keep them in sync with
> > > - * the hierarchical ones.
> > > - */
> > > - __this_cpu_add(pn->lruvec_stat_local->count[idx], x);
> > > for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
> > > atomic_long_add(x, &pi->lruvec_stat[idx]);
> > > x = 0;
> > > --
> > > 2.21.0
> > >
