> From: "Kiryl Shutsemau (Meta)" <[email protected]> > > 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.
My overall feeling is that changelogs need more human touch to be parseable. > - vm_util grows folio-order helpers: is_backed_by_folio() moves out of > split_huge_page_test.c, and a range-level form asks whether every > order-aligned window of a range is backed by one folio of exactly that > order, mapped head to tail. > > - folio_order_check validates those helpers against the kernel for every > anon THP order it supports, before any collapse test trusts them. > > - khugepaged_full_pass() drives exactly one scan pass through the sysfs > wake path, so a barrier completes on the daemon's own cadence instead of > needing a short scan_sleep_millisecs. > > - khugepaged -o <order> runs order-parameterized anon collapse cases: a > full table, only the populated window, default max_ptes_none and > max_ptes_none=0, and collapse upward from smaller large folios. Any > order up to the PMD order, where the same cases baseline the PMD path. > > - khugepaged_race races two faulters, MADV_DONTNEED, transient FOLL_PIN, > fork and mremap against one of three collapse drivers: khugepaged a pass > at a time, khugepaged free-running, or a MADV_COLLAPSE loop. Every > racing page must read as its pattern or as zero. -z collapses > hole-bearing windows; -p adds pageout and compaction, the only threads > that elevate a source folio's refcount from the reclaim side. DEBUG_VM, > page_table_check, KASAN and lockdep are the other half of the oracle. > > - A shared-source write race pins the CoW isolation contract: a co-sharer > writing throughout a collapse must not see the collapsing side's pages. > > - Four tests fail on the environment rather than on the kernel: the > collapse wait is a fixed three seconds whatever a huge page costs to > build, collapse_compound_extreme wants a 512M page from the fault path, > the shmem cases want a PMD-order page cache folio the page cache caps > below, and the swap cases fail instead of skipping without swap. Scale > the wait (a 2M PMD is unchanged) and skip the other three. This is why > the suite now runs on arm64 with 64K pages. > > Tested on mm-new (1dbd7c34bb92): > > x86-64 4K 105 pass, 3 skip, 0 fail > arm64 64K 87 pass, 7 skip, 0 fail > > The skips are structural: sub-PMD cases decline at the PMD order and the > mixed-source case at the smallest order, and on 64K pages > collapse_compound_extreme and the shmem cases skip for the reasons above. > > Kiryl Shutsemau (Meta) (16): > selftests/mm: move is_backed_by_folio() into vm_util > selftests/mm: add folio-order check for VA ranges > selftests/mm: add folio-order detection self-check > selftests/mm: add order-parameterized khugepaged collapse cases > selftests/mm: add khugepaged completion barrier helper > selftests/mm: add khugepaged race harness > selftests/mm: cover a shared-source collapse write race > selftests/mm: skip collapse_compound_extreme where the PMD is too > large > selftests/mm: skip khugepaged swap tests when no swap is configured > selftests/mm: verify synchronous khugepaged driving is attributable > selftests/mm: race-harness variant for permissive hole occupancy > selftests/mm: add memory-pressure threads to the khugepaged race > harness > selftests/mm: parameterize the mixed-source collapse case by source > order > selftests/mm: zap whole PTE tables in the khugepaged race harness > selftests/mm: scale khugepaged's collapse wait with the PMD size > selftests/mm: skip khugepaged shmem cases without a PMD page cache > folio I'd put the commits that update the existing tests at the top of the series and than add the new tests and infrastrucure. -- Sincerely yours, Mike.
