Re: [PATCH v4] mm/migrate: don't batch the TLB flush when there is nothing to batch

[Zi Yan <ziy@nvidia.com>]

On 11 Nov 2026, at 20:03, Zhiteng Qiu wrote:

> NUMA balancing migrates one folio at a time, but since commit
> 7e12beb8ca2a ("migrate_pages: batch flushing TLB") it still unmaps
> with TTU_BATCH_FLUSH like any other MIGRATE_ASYNC migration.  When the
> batch contains one small folio mapped only once, there is nothing to
> batch, and the batched unmap can leave mm->tlb_flush_batched pending,
> so the NUMA scanner's next change_pte_range() can call flush_tlb_mm()
> before changing the PTE table.  With migrations running concurrently,
> that state keeps being refilled, and on this workload the scanner
> often flushes the whole mm as it walks the PTE tables.
>
> This slows scanning and keeps mm->tlb_flush_pending raised longer.  On
> x86, a migration that unmaps a PROT_NONE PTE while
> mm->tlb_flush_pending is set gets true from pte_accessible() and needs
> a flush too; with batching, the resulting arch_tlbbatch_flush() is a
> full flush.  The two feed each other.  In one run, 33% of the
> migrations overlapped a scan pass, although passes covered only 9% of
> the time.
>
> On a 2-socket, 384-CPU x86 machine without INVLPGB, with v7.3, a
> 120 GiB anonymous working set on 4 KiB pages and 384 threads (three
> balancing on/off pairs per kernel):
>
>                                      v7.3          patched
>   throughput loss from balancing     13.0-16.5%    0.5-6.3%
>   TLB interrupts per migrated page   202-210       12.7-13.6
>   average scanner mmu_gather time    ~90 ms        ~6.7 ms
>
> The mmu_gather times were measured in separate instrumented runs.
>
> So don't ask for a batched flush when the batch holds a single small
> folio that is not mapped shared; try_to_migrate_one() then uses
> ptep_clear_flush() for each PTE it encounters.  Large folios retain
> their existing flush behavior.  migrate_pages_batch() computes the
> flags, since it knows how many folios it is migrating.  PMD-mapped
> THPs are not affected, since set_pmd_migration_entry() flushes them
> regardless of TTU_BATCH_FLUSH.
>
> Signed-off-by: Zhiteng Qiu <qiuzhiteng@hygon.cn>
> ---
> Changes in v4:
>  - Compute the try_to_migrate() flags in migrate_pages_batch() with a
>    helper and pass them to migrate_folio_unmap() instead of a bool, as
>    suggested by David.
>  - Go back to v2's check, limited to small folios, using
>    folio_maybe_mapped_shared() as Zi suggested.  folio_mapcount() would
>    stop being a good hint for large folios once try_to_unmap() batches
>    them better (David).
>  - Shorter comment (Zi).
>
> Changes in v3:
>  - Use folio_mapcount() for the check, as Ying suggested.
>  - Cut the changelog down.
>
> Changes in v2:
>  - Pass the try_to_migrate() flags directly and use single_folio in
>    the VM_WARN_ON_ONCE(), as suggested by Ying.
>  - Rewrote the changelog after looking into why b99a342d4f11 does not
>    avoid the flush here.
>
> v3: https://lore.kernel.org/r/20261109122915.2150586-1-qiuzhiteng@hygon.cn
> v2: https://lore.kernel.org/r/20261107143456.1590658-1-qiuzhiteng@hygon.cn
> v1: https://lore.kernel.org/r/20261106041628.1177963-1-qiuzhiteng@hygon.cn
>
> The table was measured on v7.3 with the same check as v4.  On mm-new
> with v4, three balancing on/off pairs gave a 1.9-2.9% loss and
> 13.2-14.6 TLB interrupts per migrated page.
>
> I re-ran the tests on v4: mm/migration selftests 6/6, LTP move_pages
> and migrate_pages 14/14, 30 minutes with compaction forced every 30 s,
> and 4.5 hours of NUMA balancing racing with fork(), MADV_DONTNEED and
> move_pages(), 4 of them on a CONFIG_DEBUG_VM kernel, all without
> warnings.  Splitting partially mapped 2 MiB THPs during NUMA migration
> still gets the batched flush (0.115-0.120 TLB interrupts per migrated
> page), and 64 KiB mTHP and a MAP_SHARED mapping shared by two processes
> behave like base.
>
> No Fixes: tag, as this is a performance trade-off that has been there
> since v6.3 and nobody has reported it as a regression.  Happy to add
> one if you think it should have one.
>
>  mm/migrate.c | 29 +++++++++++++++++++++++++----
>  1 file changed, 25 insertions(+), 4 deletions(-)
>

LGTM.

Reviewed-by: Zi Yan <ziy@nvidia.com>

Best Regards,
Yan, Zi