Re: [PATCH v2 1/3] mm: page_alloc: add a watermark check that knows the migratetype
From: Vlastimil Babka (SUSE)
Date: Fri Nov 13 2026 - 13:51:03 EST
On 11/11/26 20:49, Kiryl Shutsemau wrote:
> On Thu, Nov 05, 2026 at 05:06:39PM +0000, Harry Yoo wrote:
>> On Mon, Nov 02, 2026 at 09:39:34AM +0000, Kiryl Shutsemau (Meta) wrote:
>> > The high-order check in __zone_watermark_ok() accepts a free page of any
>> > migratetype. Under ALLOC_NOFRAGMENT, a request below pageblock_order can
>> > only take a page of its own migratetype or claim a whole free block, so
>> > the check can pass when the allocation cannot succeed.
>> >
>> > Telling those apart needs the request's migratetype, but not every check
>> > wants them told apart. A check whether an allocation can be served now
>> > has to count only the pages it can take.
>> >
>> > A check whether there is enough free memory counts every migratetype:
>> > kswapd and kcompactd balancing a zone, order-0 thresholds, the
>> > slowpath's forecast whether reclaim and compaction can still make
>> > progress. Reclaim and compaction work across migratetypes to free whole
>> > pageblocks.
>>
>> nit: perhaps we could make the changelog above bit more concise.
>>
>> I think the necessary bit here is "With defrag_mode, allocations cannot
>> steal a page from pageblocks of other migrate type, thus
>> __zone_watermark_ok() needs to take that into account."
>
>
> It is ALLOC_NOFRAGMENT, not defrag_mode: Normal zone allocations get it
> on the first attempt when DMA32 is populated, with defrag_mode off. And
> such an allocation can still claim a whole free pageblock of another
> migratetype.
>
> But fair, it can be shorter. A tighter version is below; it also says
> why kcompactd passes MIGRATE_MOVABLE.
>
>> nit: maybe let's set cc.migratetype and pass it?
>
> The value is not read there: __zone_watermark_ok_mt() uses the
> migratetype only under ALLOC_NOFRAGMENT, and kcompactd never passes it.
> kcompactd_node_suitable() makes the same call and has no compact_control
> to take it from.
>
> Thanks for the review.
>
> The new changelog:
>
> The high-order check in __zone_watermark_ok() accepts a free page of any
> migratetype. Under ALLOC_NOFRAGMENT, a request below pageblock_order can
> only take a page of its own migratetype or claim a whole free block, so
> the check can pass when the allocation cannot succeed.
>
> Only a check whether one allocation can be served now should count just
> the pages it can take.
Please try actually reading this sentence, does it make any sense? It
instantly reminded me of the meme classic: "Has anyone really been far even
as decided to use even go want to do look more like?"
> A check whether a zone has enough free memory
> serves no single allocation and counts every migratetype: reclaim and
> compaction free whole pageblocks across migratetypes.
Oh yeah always this damned style again. What did they train the models on?
> Add __zone_watermark_ok_mt(), which takes the migratetype, for the first
> kind:
>
> - the allocator fast path, zone_watermark_fast();
> - its recheck after node reclaim in get_page_from_freelist();
> - compaction's up-front check, compaction_suit_allocation_order().
>
> __zone_watermark_ok() wraps it for the second kind. kcompactd shares
> compaction_suit_allocation_order() with direct compaction and passes
> MIGRATE_MOVABLE, which is not read without ALLOC_NOFRAGMENT.
>
> Declare both in mm/internal.h: nothing outside mm uses them.
>
> Nothing reads the migratetype yet: no functional change. Preparation for
> making the high-order check honour ALLOC_NOFRAGMENT.
>
> Looks good?
Not really :(