Re: [PATCH v2 0/7] mm: batch rmap walks during large folio migration
From: Joanne Koong
Date: Thu Nov 12 2026 - 20:02:51 EST
On Wed, Nov 11, 2026 at 11:37 AM Garg, Shivank <shivankg@xxxxxxx> wrote:
>
> On Tue, 2026-11-10 at 16:08 -0800, Joanne Koong wrote:
> > On Wed, Aug 12, 2026 at 9:24 PM Shivank Garg <shivankg@xxxxxxx> wrote:
> > >
> > > Migrating a PTE-mapped large folio currently walks the folio's rmap
> > > one page at a time. It does so twice per migration:
> > > 1. try_to_migrate_one(): Replace present PTEs with migration entries
> > > 2. remove_migration_pte(): Restore migration entries to working PTEs.
> > >
> > > For a folio with N contiguous PTEs, each walk repeats page_vma_mapped_walk()
> > > and the associated operation per subpage of the folio. (256 iterations
> > > for 1M folio).
> > >
> > > This series enables both walks to process a batch of consecutive PTEs
> > > that map the same large folio and do some refactoring and clean up on
> > > the way.
> > >
> > > Performance:
> > >
> > > AMD EPYC ZEN 3, 2-socket system (1 node per socket).
> > > move_pages() migrating 1 GB from node 0 -> node 1, throughput
> > > in GB/s:
> > > batch- batch-restore
> > > folio vanilla restore-only + unmap restore% restore+unmap%
> > > 4K 2.96±0.15 3.02±0.14 2.99±0.17 +2.2% +1.0%
> > > 64K 5.99±0.10 6.12±0.08 7.65±0.16 +2.2% +27.7%
> > > 256K 6.48±0.09 6.67±0.08 9.23±0.14 +2.9% +42.4%
> > > 1M 6.76±0.10 6.98±0.10 9.74±0.19 +3.1% +44.0%
> > > 2M 10.50±0.20 10.44±0.16 10.67±0.21 -0.5% +1.6%
> > >
> > > In plain migration, the folio copy itself dominates the cost, so these rmap walk
> > > savings are only partly visible end-to-end.
> > >
> > > When the folio copy is batched-copy and offloaded to DMA engine [1], the copy is
> > > no longer the bottleneck and rmap-walk batching translates into much larger gains:
> > >
> > > This series + [1], DMA Offload on PTDMA (DCBM), 16 channels:
> > >
> > > batch-copy dcbm+restore
> > > folio vanilla -offload(dcbm) dcbm+restore +unmap (rmap batch impact)
> > > 4K 3.05±0.13 2.76±0.11 2.77±0.12 2.74±0.09
> > > 64K 5.85±0.11 10.66±0.27 11.56±1.45 17.55±0.56
> > > 256K 6.26±0.08 13.64±2.73 13.68±0.16 34.76±0.57 (2.54x)
> > > 1M 6.58±0.08 14.25±0.07 15.27±0.77 44.28±0.27 (3.11x)
> > > 2M 10.61±0.24 65.26±0.96 65.25±1.65 65.66±0.87
> > >
> >
> > Hi Shivank,
> >
> > Are you planning to send a v3 of this series? My reason for asking is
> > because batching the unmap also fixes an existing issue where when
> > try_to_migrate_one() unmaps a PTE-mapped anon mTHP one PTE at a time,
> > the folio looks partially mapped partway through and gets flagged for
> > deferred split. My mTHP underused shrinker series [1] makes this more
> > common since every underused mTHP split will now go through this path,
> > which makes this bug a blocker for the series. More context can be
> > found in [2], which was my attempt to fix this bug, but then David
> > suggested we fix it in a better way by batching, and Zi pointed me to
> > this series.
> >
> > Thanks,
> > Joanne
> >
> > [1] https://lore.kernel.org/linux-mm/20261103212004.1356327-1-joannelkoong@xxxxxxxxx/
> > [2] https://lore.kernel.org/linux-mm/20261109220958.2071687-1-joannelkoong@xxxxxxxxx/
>
>
> Glad to hear this series helps your use case.
>
> Yes, I'm working on v3. I slowed down on it because it depends on
> Optimize anonymous large folio unmapping series [1], which hasn't been
> merged yet.
Thanks for the context.
>
> If it helps, I can send v3 sooner, addressing the v2 comments, rebase
> on the current tree, and split into unmap-side and restore-side series
> for easier review.
That would be great! The unmap side is the part that fixes the
deferred split issue for me.
Thanks,
Joanne