Re: [PATCH v3 2/3] mm: page_alloc: make the high-order watermark check honour ALLOC_NOFRAGMENT

From: Kiryl Shutsemau (Meta)

Date: Thu Nov 19 2026 - 09:03:05 EST


On Wed, Nov 18, 2026 at 01:40:17PM +0530, Dev Jain wrote:
> So let us do something like: we have zone_watermark_ok_for_alloc. It will call
> zone_watermark_ok, which will only handle the watermark computation (so this is a
> refactor for the code ending at if (!order) line in current zone_watermark_ok).
> Then it will also scan the freelists for the correct order, correct migratetype etc.

A count-only zone_watermark_ok() does not work for three callers that
pass a high order: kswapd's pgdat_balanced(), kcompactd's
__compact_finished() under defrag_mode, and should_reclaim_retry(). They
ask whether the zone has a free page of that order regardless of
migratetype. They need the free list scan without the migratetype
filter, so neither of your two helpers fits them.

So the split is not count vs scan. It is the zone as a whole vs one
request, as the comment above __zone_watermark_ok() says. The _mt variant
is the one that takes the request's migratetype. Callers that only want
the count already pass order 0.

I am not particularly tied to the _mt() name, but _for_alloc() is a
mouthful.

--
Kiryl Shutsemau / Kirill A. Shutemov