On 07/08/2026 12:36 pm, Kiryl Shutsemau wrote:
> From: "Kiryl Shutsemau (Meta)" <kas@kernel.org>
>
> mTHP collapse went in for 7.2 with no functional selftest coverage. Every
> khugepaged collapse case is PMD-shaped, down to the detection: it reads
> smaps AnonHugePages, which cannot see anything below the PMD order.
>
> I am also reworking khugepaged's collapse mechanism, and that wants a suite
> worth trusting before the mechanism changes underneath it. Nothing here
> depends on the rework: everything passes on an unmodified kernel and
> documents what khugepaged already does.
>
> This version sits on top of Baolin's anon mTHP collapse selftests [1], now
> in mm-new. His mthp_khugepaged context and per-order detection are what the
> new cases need, so they go in there rather than bringing their own.
>
> Four existing cases fail on the environment rather than on the kernel:
>
> - The collapse wait is a fixed three seconds, whatever a huge page costs
> to build. Scale it with the PMD size.
>
> - collapse_compound_extreme wants a 512M folio from the fault path. Skip
> it where the PMD is that large.
>
> - The page cache caps folio order below the PMD order where a PMD is 512M,
> so nothing can hold a PMD-sized page cache folio. Skip the shmem and
> file cases there.
>
> - The swap cases ask a best-effort MADV_PAGEOUT for an exact swap count.
> Ask again before calling it a failure.
>
> This is why the suite now runs on arm64 with 64K pages.
>
> Then the new coverage:
>
> - Folio-order helpers in vm_util, and folio_order_check to validate them
> against the kernel before any collapse test trusts them.
>
> - Four order-parameterized anon collapse cases, plus mixed sources by
> source order, at every supported order by default.
>
> - A shared-source write race: a co-sharer writing throughout a collapse
> and the collapsing side must not see each other's pages.
>
> - khugepaged_full_pass(), which steps the daemon one scan at a time
> through the sysfs wake path, and khugepaged_sync_check to hold it to
> exactly one attributed collapse per step.
>
> - khugepaged_race, racing faults, MADV_DONTNEED, pins, fork and mremap
> against collapse, checked by content and by KASAN and lockdep.
>
> Tested on mm-new (8086604be973), with KASAN, lockdep, DEBUG_VM and
> page_table_check, in 16G guests:
>
> x86-64 4K 190 pass, 1 skip, 0 fail
> arm64 64K 228 pass, 5 skip, 0 fail
>
> The skips are structural: mixed sources declines at the smallest order, and
> at 64K collapse_compound_extreme and the shmem components skip as above.
I've tested them on arm64 vm and more tests are passing instead of getting
skipped. Hence:
Tested-by: Muhammad Usama Anjum <usama.anjum@arm.com>
>
> Changes since v1:
>
> - Rebased onto [1] (Baolin Wang), which this overlapped and which is in
> mm-new now, so this applies on top of it. This series' own -o mode is
> gone in favour of his context.
> - A plain ./khugepaged and ./khugepaged_race now run everything they can:
> every supported order, and every mode, occupancy and pressure axis. The
> options only narrow that down (Mike Rapoport).
> - The fixes to existing cases come first, then the infrastructure and the
> new tests, each next to what it validates (Mike Rapoport).
> - The completion barrier helper comes before the tests that use it, so no
> case carries a throwaway implementation of it (Mike Rapoport).
> - Changelogs and comments rewritten to be read rather than parsed
> (Mike Rapoport).
> - One new patch: retry the swapout, which is what made
> collapse_max_ptes_swap flaky at 64K.
> - Dropped the swap-skip patch, an equivalent has landed
> (Alexander Gordeev).
> - Review fixes: masked failures, system state left behind on bail-out
> paths, arguments taken on trust (Sashiko).
>
> [1] https://lore.kernel.org/all/cover.1785985999.git.baolin.wang@linux.alibaba.com
>
> Kiryl Shutsemau (Meta) (17):
> selftests/mm: skip collapse_compound_extreme where the PMD is too
> large
> selftests/mm: scale khugepaged's collapse wait with the PMD size
> selftests/mm: skip khugepaged page cache cases without a PMD folio
> selftests/mm: retry the swapout the khugepaged swap cases rely on
> selftests/mm: move is_backed_by_folio() into vm_util
> selftests/mm: add folio-order check for address ranges
> selftests/mm: add folio-order detection self-check
> selftests/mm: add khugepaged completion barrier helper
> selftests/mm: add order-parameterized khugepaged collapse cases
> selftests/mm: parameterize the mixed-source collapse case by source
> order
> selftests/mm: cover a shared-source collapse write race
> selftests/mm: run every supported collapse order by default
> selftests/mm: verify synchronous khugepaged driving is attributable
> selftests/mm: add khugepaged race harness
> selftests/mm: race collapse of windows with holes
> selftests/mm: add memory-pressure threads to the khugepaged race
> harness
> selftests/mm: zap whole PTE tables in the khugepaged race harness
>
> tools/testing/selftests/mm/Makefile | 3 +
> .../testing/selftests/mm/folio_order_check.c | 137 +++++
> tools/testing/selftests/mm/hmm-tests.c | 1 -
> .../testing/selftests/mm/hugepage_settings.c | 72 ++-
> .../testing/selftests/mm/hugepage_settings.h | 3 +
> tools/testing/selftests/mm/khugepaged.c | 466 +++++++++++++-
> tools/testing/selftests/mm/khugepaged_race.c | 571 ++++++++++++++++++
> .../selftests/mm/khugepaged_sync_check.c | 217 +++++++
> tools/testing/selftests/mm/run_vmtests.sh | 8 +-
> .../selftests/mm/split_huge_page_test.c | 62 --
> tools/testing/selftests/mm/vm_util.c | 145 +++++
> tools/testing/selftests/mm/vm_util.h | 10 +
> 12 files changed, 1599 insertions(+), 96 deletions(-)
> create mode 100644 tools/testing/selftests/mm/folio_order_check.c
> create mode 100644 tools/testing/selftests/mm/khugepaged_race.c
> create mode 100644 tools/testing/selftests/mm/khugepaged_sync_check.c
>
>
> base-commit: 8086604be9733378461a5534c39cb72635226b23
--
Thanks,
Usama