Re: [RFC PATCH 2/5] mm/page_alloc: use smallest fallback page first in movable allocation

[Vlastimil Babka]

On 10/14/2016 03:26 AM, Joonsoo Kim wrote:
> On Thu, Oct 13, 2016 at 11:12:10AM +0200, Vlastimil Babka wrote:
> > On 10/13/2016 10:08 AM, js1304@xxxxxxxxx wrote:
> > >From: Joonsoo Kim <iamjoonsoo.kim@xxxxxxx>
> > >
> > >When we try to find freepage in fallback buddy list, we always serach
> > >the largest one. This would help for fragmentation if we process
> > >unmovable/reclaimable allocation request because it could cause permanent
> > >fragmentation on movable pageblock and spread out such allocations would
> > >cause more fragmentation. But, movable allocation request is
> > >rather different. It would be simply freed or migrated so it doesn't
> > >contribute to fragmentation on the other pageblock. In this case, it would
> > >be better not to break the precious highest order freepage so we need to
> > >search the smallest freepage first.
> >
> > I've also pondered this, but then found a lower hanging fruit that
> > should be hopefully clear win and mitigate most cases of breaking
> > high-order pages unnecessarily:
> >
> > http://marc.info/?l=linux-mm&m=147582914330198&w=2
>
> Yes, I agree with that change. That's the similar patch what I tried
> before.
>
> "mm/page_alloc: don't break highest order freepage if steal"
> http://marc.info/?l=linux-mm&m=143011930520417&w=2

Ah, indeed, I forgot about it and had to rediscover :)

> > So I would try that first, and then test your patch on top? In your
> > patch there's a risk that we make it harder for
> > unmovable/reclaimable pageblocks to become movable again (we start
> > with the smallest page which means there's lower chance that
> > move_freepages_block() will convert more than half of the block).
>
> Indeed, but, with your "count movable pages when stealing", risk would
> disappear. :)

Hmm, but that counting is only triggered when we attempt to steal whole 
pageblock. For movable allocation, can_steal_fallback() allows that only for
(order >= pageblock_order / 2), and since your patch makes "order" as small as 
possible for movable allocations, the chances are lower?

> > And Johannes's report seems to be about a regression in exactly this
> > aspect of the heuristics.
>
> Even if your change slows down the breaking high order freepage, but,
> it would provide just a small delay to break. High order freepage
> would be broken soon and we cannot prevent to decrease high order
> freepage in the system. With my approach, high order freepage would
> stay longer time.
>
> For Johannes case, my approach doesn't aim at recovering from that
> situation. Instead, it tries to prevent such situation that
> migratetype of pageblock is changed.
>
> Thanks.