On Wed, 11 Mar 2020, Tetsuo Handa wrote:
> >>> diff --git a/mm/vmscan.c b/mm/vmscan.c
> >>> --- 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();
> >>> +
> >>
> >> Is this safe for CONFIG_PREEMPTION case? If current thread has realtime priority,
> >> can we guarantee that the OOM victim (well, the OOM reaper kernel thread rather
> >> than the OOM victim ?) gets scheduled?
> >>
> >
> > I think it's the best we can do that immediately solves the issue unless
> > you have another idea in mind?
>
> "schedule_timeout_killable(1) outside of oom_lock" or "the OOM reaper grabs oom_lock
> so that allocating threads guarantee that the OOM reaper gets scheduled" or "direct OOM
> reaping so that allocating threads guarantee that some memory is reclaimed".
>
The cond_resched() here is needed if the iteration is lengthy depending on
the number of descendant memcgs already.
schedule_timeout_killable(1) does not make any guarantees that current
will be scheduled after the victim or oom_reaper on UP systems.
If you have an alternate patch to try, we can test it. But since this
cond_resched() is needed anyway, I'm not sure it will change the result.
> >
> >>> switch (mem_cgroup_protected(target_memcg, memcg)) {
> >>> case MEMCG_PROT_MIN:
> >>> /*
> >>>
> >>
>
>
