Re: [PATCH 00/13] mm: PMD-level swap entries for anonymous THPs

[Zi Yan <ziy@nvidia.com>]

+Ying, who did the original THP swap work[1].

[1] https://lkml.org/lkml/2016/8/9/588

On 27 Apr 2026, at 6:01, Usama Arif wrote:

> When reclaim swaps out a PMD-mapped anonymous THP today, the PMD is
> split into 512 PTE-level swap entries via TTU_SPLIT_HUGE_PMD before
> unmap.
>
> This series introduces a PMD-level swap entry. The huge mapping is
> preserved across the swap round-trip, and do_huge_pmd_swap_page()
> resolves the entire 2 MB region in a single fault on swap-in,
> no khugepaged involvement is needed. swap_map metadata is identical
> either way (512 single-slot counts), so the PTE split buys nothing
> on the swap side, it is purely a page-table representation change.
>
> This work was brought about after Hugh reported that one of the
> major blockers for having lazy page table deposit is the lack of
> PMD swap entries [1]. However, this series has benefits of its
> own:
> - The huge mapping is restored on swap-in.  Today even when the
>   folio is still in swap cache as a single 2 MB folio, the swap-in
>   path installs 512 PTE mappings -- the PMD mapping is gone, the
>   freshly-materialised PTE table sticks around, and only
>   khugepaged can later collapse the range back into a THP.
>   do_huge_pmd_swap_page() reinstalls the PMD mapping directly in
>   one fault, no khugepaged involvement.
> - Memory saved per swapped-out THP *once lazy page table deposit is
>   merged* [2]. With lazy page table deposit [2], splitting a PMD into
>   512 PTE swap entries forces allocation of a 4 KB PTE table page.
>   The new path leaves the pgtable hierarchy at PMD level and avoids
>   that allocation entirely.
>   This will save memory when swapping, which is likely when there is
>   memory pressure and exactly when allocations are most likely to
>   fail.
> - Walkers (zap, mprotect, smaps, pagemap, soft-dirty, uffd-wp)
>   visit one PMD entry instead of 512 PTEs, reducing traversal
>   time and lock-hold windows.
>
> The swap entry value is identical to 512 PTE swap entries (same
> type, same starting offset), so swap_map refcounting is unchanged.
> Only the page-table representation differs; the swap slot allocator,
> swap I/O, and swap cache are untouched.  The new path falls back to
> the existing PTE-split path whenever a PMD-order resource is
> unavailable: zswap enabled, non-contiguous swap allocation
> (THP_SWPOUT_FALLBACK), PMD-order folio allocation failure on swap-in
> or fork, racing folio split, or rmap-driven split on a swapcache
> folio.  Walkers that previously assumed every non-present PMD encodes
> a PFN (migration / device_private) are taught to recognise PMD swap
> entries.
>
> Patch breakdown:
>
> The series is ordered to preserve git bisectability: every consumer
> of a PMD swap entry (split, fork, swapoff, walkers, UFFDIO_MOVE,
> swap-in fault) lands before the producer.  The swap-out path that
> actually installs PMD swap entries is the very last functional patch
> (12), so no intermediate commit can leave the kernel handling a
> PMD swap entry it does not yet understand.
>
> The first 4 patches are preparatory patches. Some of them (like
> softleaf_to_pmd() change in patch 1) are not exactly needed but its
> done to hopefully improve code quality and so that the PMD swap
> entry changes look well integrated with the rest of mm.
>
> Prep patches:
>   1. mm: add softleaf_to_pmd() and convert existing callers
>      PMD counterpart to softleaf_to_pte(); needed to construct a
>      PMD from a swap entry in later patches.
>   2. mm: extract ensure_on_mmlist() helper
>      Hoists the "register mm with swapoff" double-checked-locking
>      pattern out of try_to_unmap_one() / copy_nonpresent_pte() so
>      the PMD swap-out and PMD fork paths can reuse it without a
>      third open-coded copy.
>   3. fs/proc: use softleaf_has_pfn() in pagemap PMD walker
>      pagemap_pmd_range_thp() today calls softleaf_to_page()
>      unconditionally; a PMD swap entry has no PFN and would crash
>      it.
>   4. mm/huge_memory: move softleaf_to_folio() inside migration branch
>      change_non_present_huge_pmd() today calls softleaf_to_folio()
>      before branching on entry type, so a PMD swap entry would
>      produce a bogus folio pointer that the migration-only code
>      below would then dereference.
>
> Core patches:
>   5. PMD swap entry detection (pmd_is_swap_entry,
>      softleaf_is_valid_pmd_entry) and per-arch pmd_swp_*exclusive
>      helpers (x86/arm64/s390/riscv/loongarch).
>   6. __split_huge_pmd_locked() learns to split a PMD swap entry
>      into 512 PTE swap entries, used as the fallback when a
>      PMD-order resource is unavailable.

I was wondering how to handle insufficient memory during swap-in.
Here it is. I have not read the code, but the split should be
straightforward, since we already have a contiguous swap space at
swap-out time and the split is just to enable PTE-level swap in, right?

>   7. Fork: copy_huge_non_present_pmd() duplicates the PMD swap entry
>      in one folio_dup_swap() call, with GFP_KERNEL retry mirroring
>      copy_pte_range().
>   8. Swapoff: unuse_pmd() reads the whole 2 MB folio and reinstalls
>      the PMD; falls back to PTE-split + unuse_pte_range() on error.
>   9. Walker updates: zap_huge_pmd, change_huge_pmd,
>      change_non_present_huge_pmd, move_soft_dirty_pmd,
>      clear_soft_dirty_pmd, make_uffd_wp_pmd, smaps_pmd_entry,
>      queue_folios_pmd (mempolicy), check_pmd_state (khugepaged),
>      and the madvise_cold_or_pageout_pte_range / madvise_free_huge_pmd
>      VM_BUG_ON extensions.
>  10. UFFDIO_MOVE: move_pages_huge_pmd() learns to move a PMD swap
>      entry whole via a new move_swap_pmd() helper modeled on
>      move_swap_pte().
>  11. Swap-in: do_huge_pmd_swap_page() resolves a PMD swap fault in
>      one shot.  Handles racing splits, SWP_STABLE_WRITES read-only
>      mapping, immediate COW for write faults; falls back to PTE-split
>      on any PMD-order resource shortfall.
>  12. Swap-out: shrink_folio_list() drops TTU_SPLIT_HUGE_PMD for
>      PMD-mappable swapcache folios (when zswap is disabled), and
>      try_to_unmap_one() installs one PMD swap entry via
>      set_pmd_swap_entry() instead of splitting.
>
> Testing:
>  13. selftests/mm: 12 tests covering swap-out/in, fork, fork+COW,
>      repeated cycles, write fault, munmap, mprotect, mremap, pagemap,
>      MADV_FREE, UFFDIO_MOVE, swapoff.
>
> Making PMD swap entries work with zswap is another project on its own and
> should be in a separate follow up series.
>
> The patches are on top of mm-unstable from 23 April
> (2bcc13c29c711381d815c1ba5d5b25737400c71a).
>
> [1] https://lore.kernel.org/all/6869b7f0-84e1-fb93-03f1-9442cdfe476b@google.com/
> [2] https://lore.kernel.org/all/20260327021403.214713-1-usama.arif@linux.dev/
>
> Usama Arif (13):
>   mm: add softleaf_to_pmd() and convert existing callers
>   mm: extract ensure_on_mmlist() helper
>   fs/proc: use softleaf_has_pfn() in pagemap PMD walker
>   mm/huge_memory: move softleaf_to_folio() inside migration branch
>   mm: add PMD swap entry detection support
>   mm: add PMD swap entry splitting support
>   mm: handle PMD swap entries in fork path
>   mm: swap in PMD swap entries as whole THPs during swapoff
>   mm: handle PMD swap entries in non-present PMD walkers
>   mm: handle PMD swap entries in UFFDIO_MOVE
>   mm: handle PMD swap entry faults on swap-in
>   mm: install PMD swap entries on swap-out
>   selftests/mm: add PMD swap entry tests
>
>  arch/arm64/include/asm/pgtable.h      |   4 +
>  arch/loongarch/include/asm/pgtable.h  |  17 +
>  arch/riscv/include/asm/pgtable.h      |  15 +
>  arch/s390/include/asm/pgtable.h       |  15 +
>  arch/x86/include/asm/pgtable.h        |  15 +
>  fs/proc/task_mmu.c                    |  47 +-
>  include/linux/huge_mm.h               |  11 +
>  include/linux/leafops.h               |  44 +-
>  include/linux/swap.h                  |   4 +-
>  include/linux/vm_event_item.h         |   1 +
>  mm/hmm.c                              |   3 +-
>  mm/huge_memory.c                      | 540 +++++++++++++++++++++--
>  mm/internal.h                         |  49 +++
>  mm/khugepaged.c                       |   6 +
>  mm/madvise.c                          |   5 +-
>  mm/memory.c                           |  51 +--
>  mm/mempolicy.c                        |   2 +
>  mm/rmap.c                             |  27 +-
>  mm/swap.h                             |   7 +
>  mm/swap_state.c                       |  35 ++
>  mm/swapfile.c                         | 144 +++++-
>  mm/vmscan.c                           |  14 +-
>  mm/vmstat.c                           |   1 +
>  tools/testing/selftests/mm/Makefile   |   1 +
>  tools/testing/selftests/mm/pmd_swap.c | 607 ++++++++++++++++++++++++++
>  25 files changed, 1554 insertions(+), 111 deletions(-)
>  create mode 100644 tools/testing/selftests/mm/pmd_swap.c
>
> -- 
> 2.52.0


Best Regards,
Yan, Zi