Re: [PATCH v5 10/11] mm: install PMD swap entries on swap-out

On 06/08/2026 03:29, Luiz Capitulino wrote:
> On 2026-07-22 11:19, Usama Arif wrote:
>> Reclaim today splits a PMD-mapped anonymous THP into 512 PTE swap
>> entries before unmap, losing the huge mapping across the swap
>> round-trip and forcing khugepaged to rebuild it later. The contiguous
>> swap range was already secured when the folio was added to the swap
>> cache (a non-contiguous allocation would have split the folio earlier),
>> so the PMD can be replaced by a single PMD-level swap entry instead.
>>
>> This patch mirrors the existing PTE swap-out path at PMD granularity:
>> - shrink_folio_list() drops TTU_SPLIT_HUGE_PMD for PMD-mappable
>>    swapcache folios. zswap is handled by the PMD swap-in users: if any
>>    covered slot currently has a zswap entry, they split the PMD swap
>>    entry and fall back to the per-PTE path.
>> - try_to_unmap_one() now has a PMD branch that calls
>>    set_pmd_swap_entry() and adjusts MM_ANONPAGES / MM_SWAPENTS by
>>    HPAGE_PMD_NR before walk_done. TTU_SPLIT_HUGE_PMD remains the
>>    fallback.
>> - set_pmd_swap_entry() is the installer. Mirroring the PTE swap-out
>>    sequence at PMD granularity, it clears the present mapping (keeping
>>    the original for rollback), bumps the swap_map refcount for the
>>    folio's 512 slots, transfers the exclusive state in the swap entry,
>>    propagates the dirty bit to the folio so writeback is not lost,
>>    and installs a swap PMD that preserves the original
>>    soft-dirty / uffd-wp / exclusive bits. Any failing step rolls back
>>    the present mapping.
>>
>> The swap entry value matches what 512 PTE swap entries would encode, so
>> swap_map refcounting is unchanged: each of the 512 slots carries a
>> count of 1, released individually on later split or together on swap-in.
>>
>> Add thp_swpout_pmd to count each PMD mapping replaced by a PMD-level
>> swap entry. Unlike the folio-level thp_swpout counter, a fork-shared THP
>> can increment this counter once for each mapping; document that
>> distinction.
>>
>> Signed-off-by: Usama Arif <[email protected]>
>> ---
>>   Documentation/admin-guide/mm/transhuge.rst |  5 ++
>>   include/linux/huge_mm.h                    |  2 +
>>   include/linux/vm_event_item.h              |  1 +
>>   mm/huge_memory.c                           | 80 ++++++++++++++++++++++
>>   mm/rmap.c                                  | 19 +++++
>>   mm/vmscan.c                                |  9 ++-
>>   mm/vmstat.c                                |  1 +
>>   7 files changed, 116 insertions(+), 1 deletion(-)
>>
>> diff --git a/Documentation/admin-guide/mm/transhuge.rst b/Documentation/admin-guide/mm/transhuge.rst
>> index 16f37135ed80..b421d7982db6 100644
>> --- a/Documentation/admin-guide/mm/transhuge.rst
>> +++ b/Documentation/admin-guide/mm/transhuge.rst
>> @@ -630,6 +630,11 @@ thp_swpout
>>       is incremented every time a huge page is swapout in one
>>       piece without splitting.
>>   +thp_swpout_pmd
>> +    is incremented every time a PMD mapping is replaced by a PMD-level
>> +    swap entry. A fork-shared THP can increment this counter once for each
>> +    PMD mapping that is swapped out.
>> +
>>   thp_swpout_fallback
>>       is incremented if a huge page has to be split before swapout.
>>       Usually because failed to allocate some continuous swap space
>> diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h
>> index e7107e0991ad..41cf643a3f55 100644
>> --- a/include/linux/huge_mm.h
>> +++ b/include/linux/huge_mm.h
>> @@ -554,6 +554,8 @@ vm_fault_t do_huge_pmd_device_private(struct vm_fault *vmf);
>>     #ifdef CONFIG_THP_SWAP
>>   vm_fault_t do_huge_pmd_swap_page(struct vm_fault *vmf);
>> +int set_pmd_swap_entry(struct page_vma_mapped_walk *pvmw,
>> +               struct folio *folio);
>>   #else
>>   static inline vm_fault_t do_huge_pmd_swap_page(struct vm_fault *vmf)
>>   {
>> diff --git a/include/linux/vm_event_item.h b/include/linux/vm_event_item.h
>> index 2628ccda076a..f8fd4e13698c 100644
>> --- a/include/linux/vm_event_item.h
>> +++ b/include/linux/vm_event_item.h
>> @@ -108,6 +108,7 @@ enum vm_event_item { PGPGIN, PGPGOUT, PSWPIN, PSWPOUT,
>>           THP_ZERO_PAGE_ALLOC_FAILED,
>>           THP_SWPOUT,
>>           THP_SWPOUT_FALLBACK,
>> +        THP_SWPOUT_PMD,
>>   #endif
>>   #ifdef CONFIG_BALLOON
>>           BALLOON_INFLATE,
>> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
>> index 6ef56936ced4..c014631e5a26 100644
>> --- a/mm/huge_memory.c
>> +++ b/mm/huge_memory.c
>> @@ -5561,3 +5561,83 @@ void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
>>       trace_remove_migration_pmd(address, pmd_val(pmde));
>>   }
>>   #endif
>> +
>> +#ifdef CONFIG_THP_SWAP
>> +/**
>> + * set_pmd_swap_entry() - Replace a PMD mapping with a PMD-level swap entry.
>> + * @pvmw: Page vma mapped walk context, must have pvmw->pmd set and
>> + *        pvmw->pte NULL (i.e. PMD-mapped).
>> + * @folio: The folio being swapped out. Must be in the swap cache.
>> + *
>> + * This installs a PMD-level swap entry in place of a present PMD mapping,
>> + * avoiding the need to split the PMD into PTE-level swap entries.
>> + *
>> + * Return: 0 on success, negative error code on failure.
>> + */
>> +int set_pmd_swap_entry(struct page_vma_mapped_walk *pvmw,
>> +               struct folio *folio)
>> +{
>> +    struct vm_area_struct *vma = pvmw->vma;
>> +    struct mm_struct *mm = vma->vm_mm;
>> +    unsigned long address = pvmw->address;
>> +    unsigned long haddr = address & HPAGE_PMD_MASK;
>> +    struct page *page = folio_page(folio, 0);
>> +    bool anon_exclusive;
>> +    pmd_t pmdval;
>> +    swp_entry_t entry;
>> +    pmd_t pmdswp;
>> +
>> +    if (!(pvmw->pmd && !pvmw->pte))
>> +        return 0;
> 
> Should we call VM_WARN_ON_ONCE() and return an error instead? This
> function is only called by try_to_unmap_one() under this condition. In
> addition, returning zero would cause try_to_unmap_one() to assume that
> the PMD swap entry was installed, which is not the case.
> 
Thanks for review!

It makes sense, I will add a warning and return -EINVAL.