[PATCH 6.18.y 0/3] mm/page_alloc: fix iomap buffered write regression from 5d8edfb900d5

[Salvatore Dipietro <dipiets@amazon.it>]

Backport of 0e05a469c294 ("mm/page_alloc: avoid direct reclaim and
compaction for costly __GFP_NORETRY allocations") to 6.18.y. It fixes a
regression from 5d8edfb900d5 ("iomap: Copy larger chunks from userspace",
v6.6), which slows buffered writes under memory fragmentation.
Andrew and Vlastimil agreed it should go to stable [1].

Patches 1-2 are its prerequisites from the same mainline merge (Cc:
stable), only moved from mm/page_alloc.h to mm/internal.h. Patch 3 is
adjusted for this tree, see its changelog. Acks from Vlastimil and
Johannes on these adjustments would be appreciated.

Built, booted and tested with pgbench simple-update (PostgreSQL 18.6,
1024 clients, 96 threads, 1200s) on m8g.24xlarge (96 vCPU arm64),
12x io2 RAID0 XFS, mean of 3 runs:

  v6.18.63               87,278 TPS
  v6.18.63 + series     154,670 TPS  (+77.2%)

[1] https://lore.kernel.org/all/1fc0caa5-43b1-487f-94cf-0d1aff4c03fd@kernel.org/

Johannes Weiner (2):
  mm: page_alloc: do not give all non-blocking requests reserve access
  mm: page_alloc: remove ALLOC_NON_BLOCK from ALLOC_RESERVES

Salvatore Dipietro (1):
  mm/page_alloc: avoid direct reclaim and compaction for costly
    __GFP_NORETRY allocations

 mm/internal.h   | 14 ++++++++------
 mm/page_alloc.c | 16 ++++++++++++++--
 2 files changed, 22 insertions(+), 8 deletions(-)

-- 
2.50.1




AMAZON DEVELOPMENT CENTER ITALY SRL, viale Monte Grappa 3/5, 20124 Milano, Italia, Registro delle Imprese di Milano Monza Brianza Lodi REA n. 2504859, Capitale Sociale: 10.000 EUR i.v., Cod. Fisc. e P.IVA 10100050961, Societa con Socio Unico