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

[Joonsoo Kim]

2020년 8월 27일 (목) 오후 10:35, Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx>님이 작성:
>
> On Wed, Aug 26, 2020 at 02:12:44PM +0900, Joonsoo Kim wrote:
> > > > 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.
> > >
> > > Well it would be much simpler and won't affect most of allocations. Better than
> > > flushing pcplists IMHO.
> >
> > Hmm...Still, I'd prefer my approach.
>
> I prefer the pcp bypass approach. It's simpler and it does not incur a
> pcp drain/refill penalty.
>
> > There are two reasons. First,
> > layering problem
> > mentioned above. In rmqueue(), there is a code for MIGRATE_HIGHATOMIC.
> > As the name shows, it's for high order atomic allocation. But, after
> > skipping pcplist
> > allocation as you suggested, we could get there with order 0 request.
>
> I guess your concern is that under some circumstances that a request that
> passes a watermark check could fail due to a highatomic reserve and to
> an extent this is true. However, in that case the system is already low
> on memory depending on the allocation context, the pcp lists may get
> flushed anyway.

My concern is that non-highorder (order-0) allocation could pollute/use the
MIGRATE_HIGHATOMIC pageblock. It's reserved for highorder atomic
allocation so it's not good if an order-0 request could get there. It would
cause more fragmentation on that pageblock.

> > We can also
> > change this code, but, I'd hope to maintain current layering. Second,
> > a performance
> > reason. After the flag for nocma is up, a burst of nocma allocation
> > could come. After
> > flushing the pcplist one times, we can use the free page on the
> > pcplist as usual until
> > the context is changed.
>
> It's not guaranteed because CMA pages could be freed between the nocma save
> and restore triggering further drains due to a reschedule.  Similarly,
> a CMA allocation in parallel could refill with CMA pages on the per-cpu
> list. While both cases are unlikely, it's more unpredictable than a
> straight-forward pcp bypass.

Agreed that it's unpredictable than the pcp bypass. But, as you said,
those cases
would be rare.

> I don't really see it as a layering violation of the API because all
> order-0 pages go through the PCP lists. The fact that order-0 is serviced
> from the pcp list is an internal implementation detail, the API doesn't
> care.

What I mean is an internal implementation layering violation. We could make
a rule even in internal implementation to make code simpler and maintainable.
I guess that order-0 is serviced from the pcp list is one of those.

Anyway, although I prefer my approach, I'm okay with using pcp bypass.

Thanks.