Re: [PATCH v2] mm: synchronize with PTE table reclaim for all folio sizes

From: Barry Song

Date: Mon Nov 16 2026 - 23:54:37 EST


On Tue, Nov 17, 2026 at 11:40 AM Baolin Wang
<baolin.wang@xxxxxxxxxxxxxxxxx> wrote:
>
> Similar to commit 243cf1f67b36 ("mm: memory: fix truncation removing mapped
> folios"), the '!pmd_present()' check in page_vma_mapped_walk() can also race
> with concurrent page table reclamation via zap_empty_pte_table().
>
> page_vma_mapped_walk() is used by many functions including try_to_unmap(),
> try_to_migrate(), folio_mkclean(), pfn_mkclean_range(), remove_migration_ptes(),
> page_mapped_in_vma(), and mapping_wrprotect_range(). Except for try_to_unmap()
> and try_to_migrate(), the other callers do not perform post-walk accounting
> checks, so the race has no observable impact on them.
>
> For migration, the race can cause unnecessary retries. For reclaim via
> try_to_unmap(), the race can cause failure to reclaim a folio that is in
> fact already unmapped (as shown below), which may cause cold/hot inversion
> (e.g., in MGLRU) and increase refaults. In extreme cases, failing to
> reclaim unmapped large folios could even cause OOM in theory.
>
> CPU 0 (reclaim) CPU 1 (MADV_DONTNEED)
> -------------------------------- --------------------------------
> folio_lock(F)
> try_to_unmap(F, TTU_SYNC)
> zap_pte_range(full PMD range)
> spin_lock(ptl)
> clear PTEs
> zap_empty_pte_table()
> spin_trylock(pml)
> pmd_clear()
> spin_unlock(pml)
>
> page_vma_mapped_walk()
> observe !pmd_present(pmde)
> nr_pages < HPAGE_PMD_NR
> skip synchronization
> return false
>
> try_to_unmap() returns
> folio_mapped(F) == true
> nr_unmap_fail += folio_nr_pages(F)
> tlb_flush_mmu_tlbonly()
> tlb_flush_rmaps()
> remove F's rmaps
> spin_unlock(ptl)
>
> Commit 243cf1f67b36 ("mm: memory: fix truncation removing mapped folios")
> has already moved the PMD clear operation (via zap_empty_pte_table()) to
> after tlb_flush_rmaps(). Similarly, page_vma_mapped_walk() also needs to
> use sync_with_folio_pmd_zap() to synchronize and avoid memory ordering
> issues on weakly ordered CPUs.
>
> Regarding performance, evaluated the impact on try_to_unmap(). Mapped a file
> to generate clean large folios, then used MADV_PAGEOUT to reclaim 1G of mapped
> clean large folios, repeated 5 times, with no measurable performance impact.
> Before: 16.031 ms over 5 loops
> After: 16.303 ms over 5 loops
>
> While the race scenario described in the patch may theoretically affect
> performance, the actual impact in real workloads is likely limited I think.
> Anyway, will use memory barriers to sync for non PMD-sized folios in a
> follow-up.
>
> Lance also reported the possible racy check for small folios[1]. However,
> whether to set TTU_SYNC for small folios should be done in a separate patch
> (this is a pre-existing issue) with proper measurement (though it may be
> hard to measure the actual impact).
>
> [1] https://lore.kernel.org/all/0ddaf336-7cb7-4e83-af2c-f60d23f7aafd@xxxxxxxxx/
> Fixes: 6375e95f381e ("mm: pgtable: reclaim empty PTE page in madvise(MADV_DONTNEED)")
> Cc: stable@xxxxxxxxxxxxxxx
> Suggested-by: David Hildenbrand (Arm) <david@xxxxxxxxxx>
> Signed-off-by: Baolin Wang <baolin.wang@xxxxxxxxxxxxxxxxx>
> Acked-by: Qi Zheng <qi.zheng@xxxxxxxxx>
> Reviewed-by: Lance Yang <lance.yang@xxxxxxxxx>

Reviewed-by: Barry Song <baohua@xxxxxxxxxx>

I agree we should start by respecting PVMW_SYNC in all cases,
although there are still some open questions to address.
Whether we should extend SYNC to more folios is a separate question.