Re: [PATCH for v5.9] mm/page_alloc: handle a missing case for memalloc_nocma_{save/restore} APIs

[Joonsoo Kim]

2020년 8월 25일 (화) 오후 2:10, Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>님이 작성:
>
> On Tue, 25 Aug 2020 13:59:42 +0900 js1304@xxxxxxxxx wrote:
>
> > From: Joonsoo Kim <iamjoonsoo.kim@xxxxxxx>
> >
> > memalloc_nocma_{save/restore} APIs can be used to skip page allocation
> > on CMA area, but, there is a missing case and the page on CMA area could
> > be allocated even if APIs are used. This patch handles this case to fix
> > the potential issue.
> >
> > Missing case is an allocation from the pcplist. MIGRATE_MOVABLE pcplist
> > could have the pages on CMA area so we need to skip it if ALLOC_CMA isn't
> > specified.
> >
> > This patch implements this behaviour by checking allocated page from
> > the pcplist rather than skipping an allocation from the pcplist entirely.
> > Skipping the pcplist entirely would result in a mismatch between watermark
> > check and actual page allocation. And, it requires to break current code
> > layering that order-0 page is always handled by the pcplist. I'd prefer
> > to avoid it so this patch uses different way to skip CMA page allocation
> > from the pcplist.
> >
> > ...
> >
> > --- a/mm/page_alloc.c
> > +++ b/mm/page_alloc.c
> > @@ -3341,6 +3341,22 @@ static struct page *rmqueue_pcplist(struct zone *preferred_zone,
> >       pcp = &this_cpu_ptr(zone->pageset)->pcp;
> >       list = &pcp->lists[migratetype];
> >       page = __rmqueue_pcplist(zone,  migratetype, alloc_flags, pcp, list);
> > +#ifdef CONFIG_CMA
> > +     if (page) {
> > +             int mt = get_pcppage_migratetype(page);
> > +
> > +             /*
> > +              * pcp could have the pages on CMA area and we need to skip it
> > +              * when !ALLOC_CMA. Free all pcplist and retry allocation.
> > +              */
> > +             if (is_migrate_cma(mt) && !(alloc_flags & ALLOC_CMA)) {
> > +                     list_add(&page->lru, &pcp->lists[migratetype]);
> > +                     pcp->count++;
> > +                     free_pcppages_bulk(zone, pcp->count, pcp);
> > +                     page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
> > +             }
> > +     }
> > +#endif
> >       if (page) {
> >               __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
> >               zone_statistics(preferred_zone, zone);
>
> That's a bunch more code on a very hot path to serve an obscure feature
> which has a single obscure callsite.
>
> Can we instead put the burden on that callsite rather than upon
> everyone?  For (dumb) example, teach __gup_longterm_locked() to put the
> page back if it's CMA and go get another one?

Hmm... Unfortunately, it cannot ensure that we eventually get the non-CMA page.
I think that the only way to ensure it is to implement the
functionality here. We can
use 'unlikely' or 'static branch' to reduce the overhead for a really
rare case but
for now I have no idea how to completely remove the overhead.

Thanks.