Re: [PATCH 08/13] mm: swap in PMD swap entries as whole THPs during swapoff

From: Usama Arif

Date: Fri May 29 2026 - 11:09:17 EST




On 26/05/2026 20:44, Alexandre Ghiti wrote:
> Hi Usama,
>
> On 4/27/26 12:01, Usama Arif wrote:
>> Add unuse_pmd() and call it from unuse_pmd_range() to swap in
>> PMD-level swap entries as whole THPs during swapoff.  This mirrors
>> the existing unuse_pte_range() but operates at PMD granularity.
>>
>> If the PMD-order folio cannot be allocated, the cached folio is no
>> longer PMD-sized (e.g. split in the swap cache by
>> deferred_split_scan() or memory_failure() while the PMD swap entry
>> was installed), or the folio is not uptodate, the PMD swap entry is
>> split into PTE-level entries via __split_huge_pmd() and a non-zero
>> error is returned so unuse_pmd_range() falls through to
>> unuse_pte_range(), which handles the individual entries at order-0.
>>
>> swapin_alloc_pmd_folio() is a separate function in swap_state.c
>> as it will be reused in swapin in a later patch.
>>
>> Signed-off-by: Usama Arif <usama.arif@xxxxxxxxx>
>> ---
>>   mm/swap.h       |   7 +++
>>   mm/swap_state.c |  35 +++++++++++++
>>   mm/swapfile.c   | 137 ++++++++++++++++++++++++++++++++++++++++++++++++
>>   3 files changed, 179 insertions(+)
>>
>> diff --git a/mm/swap.h b/mm/swap.h
>> index a77016f2423b..76752df71693 100644
>> --- a/mm/swap.h
>> +++ b/mm/swap.h
>> @@ -301,6 +301,7 @@ struct folio *swap_cluster_readahead(swp_entry_t entry, gfp_t flag,
>>   struct folio *swapin_readahead(swp_entry_t entry, gfp_t flag,
>>           struct vm_fault *vmf);
>>   struct folio *swapin_folio(swp_entry_t entry, struct folio *folio);
>> +struct folio *swapin_alloc_pmd_folio(swp_entry_t entry, struct mm_struct *mm);
>>   void swap_update_readahead(struct folio *folio, struct vm_area_struct *vma,
>>                  unsigned long addr);
>>   @@ -438,6 +439,12 @@ static inline struct folio *swapin_folio(swp_entry_t entry, struct folio *folio)
>>       return NULL;
>>   }
>>   +static inline struct folio *swapin_alloc_pmd_folio(swp_entry_t entry,
>> +            struct mm_struct *mm)
>> +{
>> +    return NULL;
>> +}
>> +
>>   static inline void swap_update_readahead(struct folio *folio,
>>           struct vm_area_struct *vma, unsigned long addr)
>>   {
>> diff --git a/mm/swap_state.c b/mm/swap_state.c
>> index 1415a5c54a43..c2e8c76658f5 100644
>> --- a/mm/swap_state.c
>> +++ b/mm/swap_state.c
>> @@ -584,6 +584,41 @@ struct folio *swapin_folio(swp_entry_t entry, struct folio *folio)
>>       return swapcache;
>>   }
>>   +#ifdef CONFIG_THP_SWAP
>> +/**
>> + * swapin_alloc_pmd_folio - allocate, charge, and read a PMD-sized swap folio.
>> + * @entry: starting swap entry to swap in
>> + * @mm: mm to charge for the swap-in
>> + *
>> + * Allocate a HPAGE_PMD_ORDER folio, charge it to @mm's memcg for @entry, and
>> + * issue the swap-in via swapin_folio(). Used by callers that need to map a
>> + * PMD swap entry as a whole THP (PMD swapoff).
>> + *
>> + * Return: the swapped-in folio, or NULL on alloc/charge/swapin failure (in
>> + * which case the caller should fall back to splitting the PMD).
>> + */
>> +struct folio *swapin_alloc_pmd_folio(swp_entry_t entry, struct mm_struct *mm)
>> +{
>> +    struct folio *folio;
>> +
>> +    folio = folio_alloc(GFP_HIGHUSER_MOVABLE, HPAGE_PMD_ORDER);
>
>
> Don't you think we should back off early by passing __GFP_NORETRY? No need to keep trying under memory pressure no?

Yes, I think that makes sense, either that or GFP_TRANSHUGE_LIGHT.

I will rebase on latest, and probably use Kairuis function for allocating folios.

Thanks!

>
>
>> +    if (!folio)
>> +        return NULL;
>> +
>> +    if (mem_cgroup_swapin_charge_folio(folio, mm, GFP_KERNEL, entry)) {
>> +        folio_put(folio);
>> +        return NULL;
>> +    }
>> +
>> +    if (!swapin_folio(entry, folio)) {
>> +        folio_put(folio);
>> +        return NULL;
>> +    }
>> +
>> +    return folio;
>> +}
>> +#endif /* CONFIG_THP_SWAP */
>> +
>>   /*
>>    * Locate a page of swap in physical memory, reserving swap cache space
>>    * and reading the disk if it is not already cached.
>> diff --git a/mm/swapfile.c b/mm/swapfile.c
>> index 390f191be9a6..7256edf4ce66 100644
>> --- a/mm/swapfile.c
>> +++ b/mm/swapfile.c
>> @@ -42,6 +42,7 @@
>>   #include <linux/suspend.h>
>>   #include <linux/zswap.h>
>>   #include <linux/plist.h>
>> +#include <linux/huge_mm.h>
>>     #include <asm/tlbflush.h>
>>   #include <linux/leafops.h>
>> @@ -2519,6 +2520,130 @@ static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
>>       return 0;
>>   }
>>   +/*
>> + * unuse_pmd - Map a locked folio at PMD granularity during swapoff.
>> + *
>> + * The caller provides a locked, swapped-in folio.  Returns 0 on success
>> + * (PMD was mapped).  Returns -EAGAIN if the swap cache folio no longer
>> + * matches the entry or the PMD changed under the lock (try_to_unuse will
>> + * rescan).  Returns -EIO if the folio is not uptodate; in that case the
>> + * PMD is split so unuse_pte_range() can handle individual pages.
>> + */
>> +static int unuse_pmd(struct vm_area_struct *vma, pmd_t *pmd,
>> +             unsigned long addr, softleaf_t entry,
>> +             struct folio *folio)
>> +{
>> +    struct mm_struct *mm = vma->vm_mm;
>> +    struct page *page;
>> +    pmd_t new_pmd, old_pmd;
>> +    spinlock_t *ptl;
>> +    rmap_t rmap_flags = RMAP_NONE;
>> +    bool exclusive;
>> +
>> +    if (unlikely(!folio_matches_swap_entry(folio, entry)))
>> +        return -EAGAIN;
>> +
>> +    if (unlikely(!folio_test_uptodate(folio))) {
>> +        __split_huge_pmd(vma, pmd, addr, false);
>> +        return -EIO;
>> +    }
>> +
>> +    page = folio_page(folio, 0);
>> +
>> +    ptl = pmd_lock(mm, pmd);
>> +    old_pmd = pmdp_get(pmd);
>> +
>> +    if (!pmd_is_swap_entry(old_pmd) ||
>> +        softleaf_from_pmd(old_pmd).val != entry.val) {
>> +        spin_unlock(ptl);
>> +        return -EAGAIN;
>> +    }
>> +
>> +    exclusive = pmd_swp_exclusive(old_pmd);
>> +
>> +    /*
>> +     * Some architectures may have to restore extra metadata to the folio
>> +     * when reading from swap. This metadata may be indexed by swap entry
>> +     * so this must be called before folio_put_swap().
>> +     */
>> +    arch_swap_restore(folio_swap(entry, folio), folio);
>> +
>> +    add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
>> +    add_mm_counter(mm, MM_SWAPENTS, -HPAGE_PMD_NR);
>> +
>> +    new_pmd = folio_mk_pmd(folio, vma->vm_page_prot);
>> +    new_pmd = pmd_mkold(new_pmd);
>> +    if (pmd_swp_soft_dirty(old_pmd))
>> +        new_pmd = pmd_mksoft_dirty(new_pmd);
>> +    if (pmd_swp_uffd_wp(old_pmd))
>> +        new_pmd = pmd_mkuffd_wp(new_pmd);
>> +
>> +    if (exclusive)
>> +        rmap_flags |= RMAP_EXCLUSIVE;
>> +
>> +    folio_get(folio);
>> +    if (!folio_test_anon(folio))
>> +        folio_add_new_anon_rmap(folio, vma, addr, rmap_flags);
>> +    else
>> +        folio_add_anon_rmap_pmd(folio, page, vma, addr, rmap_flags);
>> +
>> +    set_pmd_at(mm, addr, pmd, new_pmd);
>> +    folio_put_swap(folio, NULL);
>> +
>> +    spin_unlock(ptl);
>> +
>> +    folio_free_swap(folio);
>> +    return 0;
>> +}
>> +
>> +/*
>> + * Try to swap in a PMD swap entry as a whole THP.  Returns 0 on success.
>> + * Returns -ENOMEM if the PMD-order folio could not be allocated/charged,
>> + * -EIO if swap-in failed, or -EAGAIN if the cached folio is no longer
>> + * PMD-sized; in all of these the PMD is split so the caller can fall
>> + * back to unuse_pte_range().  Otherwise propagates the error from
>> + * unuse_pmd().
>> + */
>> +static int unuse_pmd_entry(struct vm_area_struct *vma, pmd_t *pmd,
>> +               unsigned long addr, softleaf_t entry)
>> +{
>> +    struct folio *folio;
>> +    int ret;
>> +
>> +    folio = swap_cache_get_folio(entry);
>> +    if (!folio) {
>> +        folio = swapin_alloc_pmd_folio(entry, vma->vm_mm);
>> +        if (!folio) {
>> +            ret = -ENOMEM;
>> +            goto split_fallback;
>> +        }
>> +    }
>> +
>> +    folio_lock(folio);
>> +    folio_wait_writeback(folio);
>> +    /*
>> +     * If the cached folio is no longer PMD-sized (e.g. split in the
>> +     * swap cache by deferred_split_scan() or memory_failure() while
>> +     * the PMD swap entry was installed), the PMD swap entry no longer
>> +     * maps a single contiguous folio.  Split the PMD swap entry so
>> +     * unuse_pte_range() can swap the per-slot folios in individually.
>> +     */
>> +    if (folio_nr_pages(folio) != HPAGE_PMD_NR) {
>> +        folio_unlock(folio);
>> +        folio_put(folio);
>> +        ret = -EAGAIN;
>> +        goto split_fallback;
>> +    }
>> +    ret = unuse_pmd(vma, pmd, addr, entry, folio);
>> +    folio_unlock(folio);
>> +    folio_put(folio);
>> +    return ret;
>> +
>> +split_fallback:
>> +    __split_huge_pmd(vma, pmd, addr, false);
>> +    return ret;
>> +}
>> +
>>   static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
>>                   unsigned long addr, unsigned long end,
>>                   unsigned int type)
>> @@ -2531,6 +2656,18 @@ static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
>>       do {
>>           cond_resched();
>>           next = pmd_addr_end(addr, end);
>> +
>> +        pmd_t pmdval = pmdp_get(pmd);
>> +
>> +        if (pmd_is_swap_entry(pmdval)) {
>> +            softleaf_t sl = softleaf_from_pmd(pmdval);
>> +
>> +            if (swp_type(sl) == type) {
>> +                if (!unuse_pmd_entry(vma, pmd, addr, sl))
>> +                    continue;
>> +            }
>> +        }
>> +
>>           ret = unuse_pte_range(vma, pmd, addr, next, type);
>>           if (ret)
>>               return ret;