Re: [PATCH 0/6] mm: batch rmap walks during large folio migration

[David Hildenbrand (Arm)]

On 7/12/26 09:38, Shivank Garg 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        unmap          restore%  restore+unmap%
> 4K        2.98±0.09      2.95±0.07      3.06±0.05        -1.0%    +2.6%
> 64K       6.16±0.09      6.24±0.08      7.80±0.08        +1.3%   +26.6%
> 256K      6.85±0.06      6.85±0.09      9.54±0.19        +0.0%   +39.3%
> 1M        7.06±0.09      7.17±0.09      9.89±0.06        +1.5%   +40.1%
> 2M       10.97±0.17     10.67±0.13     10.69±0.15        -2.7%    -2.6%
> 
> In plain migration, the folio copy itself dominate 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 bottlneck and rmap-walk batching translates into much larger gains:
> 
> This series + [1], DMA Offload on PTDMA (DCBM), 16 channels:
> 
> 		     batch-copy
> folio   vanilla      -offload(dcbm)  dcbm+restore   dcbm+restore+unmap
> 4K      3.36 ± 0.10   4.16 ± 0.18     4.10 ± 0.14    4.12 ± 0.13
> 64K     6.16 ± 0.47   9.63 ± 0.23    10.00 ± 0.33   14.63 ± 0.14
> 256K    6.70 ± 0.09  12.98 ± 0.18    13.52 ± 0.15   32.30 ± 0.65    (2.38x)
> 1M      6.99 ± 0.02  14.32 ± 0.67    14.84 ± 0.17   38.97 ± 1.03    (2.62x)
> 2M     10.74 ± 0.29  64.25 ± 1.48    65.36 ± 1.00   64.97 ± 1.75
> 
> [1] https://lore.kernel.org/linux-mm/20260630-shivank-batch-migrate-offload-v6-0-da95d7e8b8a2@xxxxxxx
> 
> Applies cleanly on mm-new 10-07-26 (61cccb8363f).
> 
> Signed-off-by: Shivank Garg <shivankg@xxxxxxx>
> ---
> Dev Jain (1):
>       mm/rmap: Add batched version of folio_try_share_anon_rmap_pte

This would ideally be based on Dev's series instead if including a patch from
his series.

-- 
Cheers,

David