]> git.ipfire.org Git - thirdparty/linux.git/commitdiff
mm: huge_memory: add folio_mark_accessed() when zapping file THP
authorBaolin Wang <baolin.wang@linux.alibaba.com>
Wed, 9 Apr 2025 09:38:58 +0000 (17:38 +0800)
committerAndrew Morton <akpm@linux-foundation.org>
Mon, 12 May 2025 00:48:26 +0000 (17:48 -0700)
When investigating performance issues during file folio unmap, I noticed
some behavioral differences in handling non-PMD-sized folios and PMD-sized
folios.  For non-PMD-sized file folios, it will call folio_mark_accessed()
to mark the folio as having seen activity, but this is not done for
PMD-sized folios.

This might not cause obvious issues, but a potential problem could be
that, it might lead to reclaim of hot file folios under memory pressure,
as quoted from Johannes:

: Sometimes file contents are only accessed through relatively short-lived
: mappings. But they can nevertheless be accessed a lot and be hot. It's
: important to not lose that information on unmap, and end up kicking out a
: frequently used cache page.

Therefore, we should also add folio_mark_accessed() for PMD-sized file
folios when unmapping.

[baolin.wang@linux.alibaba.com: add comment]
Link: https://lkml.kernel.org/r/23fdc11d-e983-4627-89a8-79e9ecf9a45a@linux.alibaba.com
Link: https://lkml.kernel.org/r/fc117f60d7b686f87067f36a0ef7cdbc3a78109c.1744190345.git.baolin.wang@linux.alibaba.com
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Zi Yan <ziy@nvidia.com>
Acked-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Cc: Barry Song <21cnbao@gmail.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
mm/huge_memory.c

index 1cd975503131ad83af3fda4a3069b34380c5ae71..5576a08a593dc6111a7e64921d7404932fb4e601 100644 (file)
@@ -2259,6 +2259,14 @@ int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
                                zap_deposited_table(tlb->mm, pmd);
                        add_mm_counter(tlb->mm, mm_counter_file(folio),
                                       -HPAGE_PMD_NR);
+
+                       /*
+                        * Use flush_needed to indicate whether the PMD entry
+                        * is present, instead of checking pmd_present() again.
+                        */
+                       if (flush_needed && pmd_young(orig_pmd) &&
+                           likely(vma_has_recency(vma)))
+                               folio_mark_accessed(folio);
                }
 
                spin_unlock(ptl);