On Thu, May 4, 2023 at 2:23 PM Zhaoyang Huang <huangzhaoyang@xxxxxxxxx> wrote:
>
> On Thu, May 4, 2023 at 12:30 AM Roman Gushchin <roman.gushchin@xxxxxxxxx> wrote:
> >
> > On Wed, May 03, 2023 at 03:58:21PM +0800, Zhaoyang Huang wrote:
> > > On Wed, May 3, 2023 at 6:01 AM Roman Gushchin <roman.gushchin@xxxxxxxxx> wrote:
> > > >
> > > > On Tue, May 02, 2023 at 12:12:28PM +0000, 黄朝阳 (Zhaoyang Huang) wrote:
> > > > > > Hi Zhaoyang!
> > > > > >
> > > > > > On Fri, Apr 28, 2023 at 07:00:41PM +0800, zhaoyang.huang wrote:
> > > > > > > From: Zhaoyang Huang <zhaoyang.huang@xxxxxxxxxx>
> > > > > > >
> > > > > > > Please be notice bellowing typical scenario that commit 168676649
> > > > > > > introduce, that is, 12MB free cma pages 'help' GFP_MOVABLE to keep
> > > > > > > draining/fragmenting U&R page blocks until they shrink to 12MB without
> > > > > > > enter slowpath which against current reclaiming policy. This commit change
> > > > > > the criteria from hard coded '1/2'
> > > > > > > to watermark check which leave U&R free pages stay around WMARK_LOW
> > > > > > > when being fallback.
> > > > > >
> > > > > > Can you, please, explain the problem you're solving in more details?
> > > > > I am trying to solve a OOM problem caused by slab allocation fail as all free pages are MIGRATE_CMA by applying 168676649, which could help to reduce the fault ration from 12/20 to 2/20. I noticed it introduce the phenomenon which I describe above.
> > > > > >
> > > > > > If I understand your code correctly, you're effectively reducing the use of cma
> > > > > > areas for movable allocations. Why it's good?
> > > > > Not exactly. In fact, this commit lead to the use of cma early than it is now, which could help to protect U&R be 'stolen' by GFP_MOVABLE. Imagine this scenario, 30MB total free pages composed of 10MB CMA and 20MB U&R, while zone's watermark low is 25MB. An GFP_MOVABLE allocation can keep stealing U&R pages(don't meet 1/2 criteria) without enter slowpath(zone_watermark_ok(WMARK_LOW) is true) until they shrink to 15MB. In my opinion, it makes more sense to have CMA take its duty to help movable allocation when U&R lower to certain zone's watermark instead of when their size become smaller than CMA.
> > > > > > Also, this is a hot path, please, make sure you're not adding much overhead.
> > > > > I would like to take more thought.
> > > >
> > > > Got it, thank you for the explanation!
> > > >
> > > > How about the following approach (completely untested)?
> > > >
> > > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > > > index 6da423ec356f..4b50f497c09d 100644
> > > > --- a/mm/page_alloc.c
> > > > +++ b/mm/page_alloc.c
> > > > @@ -2279,12 +2279,13 @@ __rmqueue(struct zone *zone, unsigned int order, int migratetype,
> > > > if (IS_ENABLED(CONFIG_CMA)) {
> > > > /*
> > > > * Balance movable allocations between regular and CMA areas by
> > > > - * allocating from CMA when over half of the zone's free memory
> > > > - * is in the CMA area.
> > > > + * allocating from CMA when over half of the zone's easily
> > > > + * available free memory is in the CMA area.
> > > > */
> > > > if (alloc_flags & ALLOC_CMA &&
> > > > zone_page_state(zone, NR_FREE_CMA_PAGES) >
> > > > - zone_page_state(zone, NR_FREE_PAGES) / 2) {
> > > > + (zone_page_state(zone, NR_FREE_PAGES) -
> > > > + zone->_watermark[WMARK_LOW]) / 2) {
> > > IMO, we should focus on non-cma area which trigger use of cma when
> > > they are lower than corresponding watermark(there is still
> > > WMARK_MIN/HIGH to deal with within slowpath)
> > > > page = __rmqueue_cma_fallback(zone, order);
> > > > if (page)
> > > > return page;
> > > >
> > > > Basically the idea is to keep free space equally split between cma and non-cma areas.
> > > > Will it work for you?
> > > I don't think it makes sense to 'equally split' cma and non-cma areas
> > > over free space while cma could occupy various proportions in a single
> > > zone. This fixed 1/2 could lead to different situation on 20% or 50%
> > > cma occupation.
> >
> > Can you then, please, explain in details what you're proposing instead
> > and why it's better (not only in your case, but generally as well)?
> > For the context, my original usecase was cma size under 10G and
> > the total memory size between 32G and 128G.
>
> Let us look at the series of scenarios below with
> WMARK_LOW=25MB,WMARK_MIN=5MB(managed pages 1.9GB). We can know that
> 1. optimized 1/2 method start to use CMA since D which actually has
> caused U&R lower than WMARK_LOW (this could be deemed as against
> current memory policy, that is, U&R should either keeping stay around
> WATERMARK_LOW or enter slowpath to do reclaiming)
> 2. it keep using CMA as long as free pages keep shrinking (this is
> against the original desire of balance cma & none-cma area)
>
> free_cma/free_pages(MB) A(12/30) B(10/25) C(8/25)
> D(8/20) E(8/16) F(5/12) F(5/10) G(4/8)
> optimized 1/2 N N
> N Y Y Y
> Y Y
> !zone_watermark_ok Y Y
> Y N N N Y
> Y
sorry for the broken graph format of previous reply, repost it here.N
stands for not use CMA while Y stands for using
free_cma/free_pages(MB) A(12/30) B(10/25) C(8/25)
optimized 1/2 N N
N
!zone_watermark_ok Y Y Y
D(8/20) E(8/16) F(5/12) F(5/10) G(4/8)
Y Y Y Y Y
N N N Y Y
> >
> > Looking at your original patch, I see that __if_use_cma_first() always returns
> > false if zone_watermark_ok() returns true. It worries me.
> ok. we could deal with the scenario here when free pages is higher
> than WMARK_HIGH
> >
> > Thanks!
