Re: [PATCH v1] mm/rmap: fix migration entries being treated as partial unmaps

["David Hildenbrand (Arm)" <david@kernel.org>]

On 11/9/26 23:09, Joanne Koong wrote:
> When try_to_migrate_one() unmaps a large anonymous folio that is mapped by
> PTEs, it replaces the PTEs with migration entries one at a time and drops
> the rmap for each page with folio_remove_rmap_pte(). After the first PTE
> is replaced, __folio_remove_rmap() sees a folio that is no longer fully
> mapped but still has pages mapped, and calls
> deferred_split_folio(folio, true). That marks the folio
> PG_partially_mapped, bumps split_deferred and nr_anon_partially_mapped,
> and puts it on the deferred split queue, even though nothing was actually
> unmapped. The migration entries are turned back into mappings once the
> split or migration is done.
> 
> This happens whenever a fully mapped, PTE-mapped anon large folio is
> migrated or split. Its effects:
> 
>  - Migrated folios can later be split for no reason. migrate_folio_move()
>    copies the source folio's deferred split state to the destination, but
>    reads it after the unmap, when PG_partially_mapped has already been
>    wrongly set. So a fully mapped mTHP ends up queued as partially mapped
>    after migration, and the shrinker can later split it without checking
>    whether it is underused.
> 
>  - split_deferred is overcounted. Every such split increments it, so the
>    stat counts folios that were never partially mapped.
> 
>  - A folio whose split fails can later be split for no reason. If the
>    split fails at folio_ref_freeze(), the folio is remapped but stays on
>    the deferred split queue marked partially mapped. The deferred split
>    shrinker then splits it without checking whether it is underused,
>    breaking up a folio that is fully mapped and possibly fully in use.

This is a similar problem I raised in last year's LSF/MM:

https://share.daveh.de/s/bQsDpD7xigJ6wD8/download?path=&files=LSF_MM%2726_+Deferred_folio_splitting.pdf

"
2. THPs erroneously detected as partially-mapped
 * Failing migration/split/eviction after stepwise unmap
 * Remapping the folio will not fix this up
"

The news here is that we now copy the state to the new folio and can
consequently split the new folio.


> 
> Mappings replaced by migration entries are not really unmapped, so don't
> queue the folio for deferred split when try_to_migrate_one() installs a
> migration entry. Pass a flag down to __folio_remove_rmap() and have
> try_to_migrate_one() clear it for those PTEs only. PTEs that it replaces
> with a hwpoison entry or discards because they are unused are really
> unmapped and keep the current behaviour, as do all other callers.
> 
> One consequence is that if part of the folio is really unmapped while the
> migration entries are in place, zapping the migration entries does not go
> through the rmap, so the folio is not marked partially mapped afterwards.
> For a split this does not matter, as the folio is split anyway. For a
> migrated folio it means the destination is not queued until a later unmap
> notices it, which seems preferable to queueing every migrated folio.
> 
> Fixes: 8422acdc97ed ("mm: introduce a pageflag for partially mapped folios")
> Fixes: a2e0c0668a34 ("mm: migrate: requeue destination folio on deferred split queue")
> Reported-by: Sashiko <sashiko-bot@kernel.org>
> Closes: https://lore.kernel.org/all/sashiko-outbox-198104@kernel.org/
> Signed-off-by: Joanne Koong <joannelkoong@gmail.com>
> ---
> This bug was found by Sashiko in its review of the mTHP shrinker series
> (https://lore.kernel.org/linux-mm/sashiko-outbox-198104@kernel.org/). The
> migration / rmap code is still new to me. This patch is a fix with the help of
> AI, so I'd appreciate a careful look from people who know this code better.

Don't we have the same problem for try_to_unmap() in some remaining cases?

See below.

> 
> This was tested in an x86 QEMU guest with 64K mTHP. Splitting 64 fully mapped
> folios through the split_huge_pages debugfs file no longer increases
> split_deferred (was +64), and migrating 64 to another node no longer leaves
> nr_anon_partially_mapped up (was +64). split_huge_page_test, migration, and
> khugepaged selftests pass.
> 
>  mm/rmap.c | 29 +++++++++++++++++++++++------
>  1 file changed, 23 insertions(+), 6 deletions(-)
> 
> diff --git a/mm/rmap.c b/mm/rmap.c
> index fa34e1ccbd8f..e5d17481c0c0 100644
> --- a/mm/rmap.c
> +++ b/mm/rmap.c
> @@ -1799,7 +1799,7 @@ void folio_add_file_rmap_pud(struct folio *folio, struct page *page,
>  
>  static __always_inline void __folio_remove_rmap(struct folio *folio,
>  		struct page *page, int nr_pages, struct vm_area_struct *vma,
> -		enum pgtable_level level)
> +		enum pgtable_level level, bool may_defer_split)
>  {
>  	atomic_t *mapped = &folio->_nr_pages_mapped;
>  	int last = 0, nr = 0, nr_pmdmapped = 0;
> @@ -1888,8 +1888,13 @@ static __always_inline void __folio_remove_rmap(struct folio *folio,
>  	 *
>  	 * Device private folios do not support deferred splitting and
>  	 * shrinker based scanning of the folios to free.
> +	 *
> +	 * Skip this if the caller is only replacing the mappings with
> +	 * migration entries. The folio is not really becoming partially
> +	 * mapped. Once the split or migration finishes, every migration
> +	 * entry becomes a mapping again.
>  	 */
> -	if (partially_mapped && folio_test_anon(folio) &&
> +	if (may_defer_split && partially_mapped && folio_test_anon(folio) &&
>  	    !folio_test_partially_mapped(folio) &&
>  	    !folio_is_device_private(folio))
>  		deferred_split_folio(folio, true);
> @@ -1921,7 +1926,7 @@ static __always_inline void __folio_remove_rmap(struct folio *folio,
>  void folio_remove_rmap_ptes(struct folio *folio, struct page *page,
>  		int nr_pages, struct vm_area_struct *vma)
>  {
> -	__folio_remove_rmap(folio, page, nr_pages, vma, PGTABLE_LEVEL_PTE);
> +	__folio_remove_rmap(folio, page, nr_pages, vma, PGTABLE_LEVEL_PTE, true);
>  }
>  
>  /**
> @@ -1938,7 +1943,8 @@ void folio_remove_rmap_pmd(struct folio *folio, struct page *page,
>  		struct vm_area_struct *vma)
>  {
>  #ifdef CONFIG_TRANSPARENT_HUGEPAGE
> -	__folio_remove_rmap(folio, page, HPAGE_PMD_NR, vma, PGTABLE_LEVEL_PMD);
> +	__folio_remove_rmap(folio, page, HPAGE_PMD_NR, vma, PGTABLE_LEVEL_PMD,
> +			    true);
>  #else
>  	WARN_ON_ONCE(true);
>  #endif
> @@ -1959,7 +1965,8 @@ void folio_remove_rmap_pud(struct folio *folio, struct page *page,
>  {
>  #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
>  	defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
> -	__folio_remove_rmap(folio, page, HPAGE_PUD_NR, vma, PGTABLE_LEVEL_PUD);
> +	__folio_remove_rmap(folio, page, HPAGE_PUD_NR, vma, PGTABLE_LEVEL_PUD,
> +			    true);
>  #else
>  	WARN_ON_ONCE(true);
>  #endif
> @@ -2516,6 +2523,8 @@ static bool try_to_migrate_one(struct folio *folio, struct vm_area_struct *vma,
>  	mmu_notifier_invalidate_range_start(&range);
>  
>  	while (page_vma_mapped_walk(&pvmw)) {
> +		bool may_defer_split = true;
> +
>  		/* PMD-mapped THP migration entry */
>  		if (!pvmw.pte) {
>  			__maybe_unused unsigned long pfn;
> @@ -2772,12 +2781,20 @@ static bool try_to_migrate_one(struct folio *folio, struct vm_area_struct *vma,
>  			 * No need to invalidate here it will synchronize on
>  			 * against the special swap migration pte.
>  			 */
> +
> +			/*
> +			 * The folio only looks partially unmapped while this PTE
> +			 * holds a migration entry, which is later turned back
> +			 * into a mapping. Don't queue it for deferred split.
> +			 */
> +			may_defer_split = false;
>  		}
>  
>  		if (unlikely(folio_test_hugetlb(folio)))
>  			hugetlb_remove_rmap(folio);
>  		else
> -			folio_remove_rmap_pte(folio, subpage, vma);
> +			__folio_remove_rmap(folio, subpage, 1, vma,
> +					    PGTABLE_LEVEL_PTE, may_defer_split);
We will not add such magical boolean parameters. I have a patch to just pass the
rmap flags for a different purpose.

But the whole thing is rather ugly ...


You should probably look at

commit 354dffd29575cdf13154e8fb787322354aa9efc4
Author: Barry Song <baohua@kernel.org>
Date:   Fri Feb 14 22:30:14 2025 +1300

    mm: support batched unmap for lazyfree large folios during reclamation

    Currently, the PTEs and rmap of a large folio are removed one at a time.
    This is not only slow but also causes the large folio to be unnecessarily
    added to deferred_split, which can lead to races between the
    deferred_split shrinker callback and memory reclamation.  This patch
    releases all PTEs and rmap entries in a batch.  Currently, it only handles
    lazyfree large folios.

Batching resolves the problem in the common case. For the remaining parts, we
should likely just check in the deferred split handler whether actually
partially mapped to then clear the flag if not.

-- 
Cheers,

David