On Tue, Aug 31, 2021 at 11:48:28AM -0400, Rik van Riel wrote:
> On Tue, 2021-08-31 at 11:59 +0200, Michal Hocko wrote:
> > On Mon 30-08-21 16:48:03, Johannes Weiner wrote:
> >
> >
> > > Or go back to not taking the branch in the first place when there
> > > is
> > > no protection in effect...
> > >
> > > diff --git a/mm/vmscan.c b/mm/vmscan.c
> > > index 6247f6f4469a..9c200bb3ae51 100644
> > > --- a/mm/vmscan.c
> > > +++ b/mm/vmscan.c
> > > @@ -2547,7 +2547,7 @@ static void get_scan_count(struct lruvec
> > > *lruvec, struct scan_control *sc,
> > > mem_cgroup_protection(sc->target_mem_cgroup, memcg,
> > > &min, &low);
> > >
> > > - if (min || low) {
> > > + if (min || (!sc->memcg_low_reclaim && low)) {
> > > /*
> > > * Scale a cgroup's reclaim pressure by
> > > proportioning
> > > * its current usage to its memory.low or
> > > memory.min
> >
> > This is slightly more complex to read but it is also better than +1
> > trick.
>
> We could also fold it into the helper function, with
> mem_cgroup_protection deciding whether to use low or
> min for the protection limit, and then we key the rest
> of our decisions off that.
>
> Wait a minute, that's pretty much what mem_cgroup_protection
> looked like before f56ce412a59d ("mm: memcontrol: fix occasional
> OOMs due to proportional memory.low reclaim")
>
> Now I'm confused how that changeset works.
>
> Before f56ce412a59d, mem_cgroup_protection would return
> memcg->memory.emin if sc->memcg_low_reclaim is true, and
> memcg->memory.elow when not.
>
> After f56ce412a59d, we still do the same thing. We just
> also set sc->memcg_low_skipped so we know to come back
> if we could not hit our target without skipping groups
> with memory.low protection...
Yeah, I just bubbled the sc->memcg_low_reclaim test up into vmscan.c
so we can modify sc->memcg_low_skipped based on it. Because
scan_control is vmscan.c-scope and I tried to retain that; and avoid
doing things like passing &sc->memcg_low_skipped into memcg code.
