On Tue, 10 Mar 2020, Andrew Morton wrote:
> > --- a/mm/vmscan.c
> > +++ b/mm/vmscan.c
> > @@ -2637,6 +2637,8 @@ static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
> > unsigned long reclaimed;
> > unsigned long scanned;
> >
> > + cond_resched();
> > +
> > switch (mem_cgroup_protected(target_memcg, memcg)) {
> > case MEMCG_PROT_MIN:
> > /*
>
>
> Obviously better, but this will still spin wheels until this tasks's
> timeslice expires, and we might want to do something to help ensure
> that the victim runs next (or soon)?
>
We used to have a schedule_timeout_killable(1) to address exactly that
scenario but it was removed in 4.19:
commit 9bfe5ded054b8e28a94c78580f233d6879a00146
Author: Michal Hocko <mhocko@xxxxxxxx>
Date: Fri Aug 17 15:49:04 2018 -0700
mm, oom: remove sleep from under oom_lock
This is why we don't see this issue on 4.14 guests but we do on 4.19. I
had assumed the issue Tetsuo reported that resulted in that patch was
still an issue and I preferred to fix the weird UP issue by adding a
cond_resched() that is likely needed for the iteration in
shrink_node_memcg() anyway. Do we care to optimize for UP systems
encountering memcg oom kills? Eh, maybe, but I'm not very interested in
opening up a centithread about this.
> (And why is shrink_node_memcgs compiled in when CONFIG_MEMCG=n?)
>
This guest does have CONFIG_MEMCG enabled, it's a memcg oom condition.
But unrelated to this patch, I think it's just a weird naming for it. The
do-while loop in shrink_node_memcgs() actually uses memcg = NULL for the
non-memcg case and is responsible for calling into page and slab reclaim.
