[PATCH v4 0/4] make unused huge shrinker memcg aware

From: Qi Zheng <[email protected]>

Changes in v4:
 - add [PATCH v4 2/4] to make obj_cgroup_memcg() handle NULL objcg
 - store obj_cgroup instead of mem_cgroup in shmem_inode_info to avoid
   pinning a dying memcg through a long-lived CSS reference
   (pointed by sashiko)
 - fix is_shmem_unused_huge_match() to always check the NUMA node for
   shrinker reclaim, not only for non-root memcg reclaim
   (pointed by sashiko)
 - collect Reviewed-by
 - rebase onto the next-20260814

Note: [PATCH v4 1/4] should ideally be folded into commit 0ef8faff490be
("fs: push nr_cached_objects memcg gating into individual filesystems") in
linux-next.

Changes in v3:
 - add a fix patch to fix missed removal of super_fs_objects_eligible()
 - move the original shrinklist addition logic after all checks are completed,
   and split it into a separate patch. (suggested by Baolin)
 - simplify the shmem_unused_huge_requeue() (suggested by Baolin)
 - keep the move_back label in shmem_unused_huge_shrink() (suggested by Baolin)
 - rebase onto the next-20260731

Changes in v2:
 - temporarily add the dependent patch from Usama to the series for review
   convenience
 - remove shrinklist_scan and shrinklist_isolated from struct shmem_inode_info,
   and re-implement the logic by resuing the same info->shrinklist
   (suggested by Baolin)
 - add more comments (suggested by Andrew)
 - fix missing initialization of info->shrinklist_memcg (pointed by sashiko)
 - rebase onto the next-20260717

Hi all,

The shmem unused huge shrinker maintains a per-superblock list of inodes
whose tail huge folio extends beyond i_size.  Because this list is not
memcg aware, reclaim triggered by memcg A can scan inodes across the
entire superblock and split huge folios charged to unrelated memcg B,
causing unexpected impact on it.

In the worst case, memcg A has no reclaimable shmem at all, making the
reclaim entirely useless and incurring unnecessary latency.  We observed
this in production, where page lock contention during split caused
multi-hundred-millisecond stalls:

  tid 11340 comm scanner locked a page for 182264 us! kstack:
          unlock_page+1
          split_huge_page_to_list+3135
          shmem_unused_huge_shrink+767
          super_cache_scan+329
          do_shrink_slab+291
          shrink_slab+533
          shrink_node+400
          do_try_to_free_pages+206
          try_to_free_mem_cgroup_pages+262
          try_charge_memcg+591
          mem_cgroup_charge+136
          __handle_mm_fault+2431
          handle_mm_fault+194
          do_user_addr_fault+462
          __do_page_fault+176
          do_page_fault+48
          page_fault+62

Usama's recent patch [1] prevents the shmem unused shrinker from being
invoked during memcg-level reclaim altogether, but this is overly
conservative: we can do better by reclaiming only the shmem charged to
the reclaiming memcg.

This series converts the shrinker list to a memcg-aware list_lru, so
that non-root memcg reclaim walks only candidates charged to the
reclaiming memcg.  Global reclaim, root memcg reclaim and shmem quota
reclaim retain their existing global semantics.

To avoid pinning a dying memcg through a long-lived CSS reference, each
inode stores an obj_cgroup reference instead of a mem_cgroup reference.
The list_lru add/delete paths resolve the current memcg from the objcg
under RCU, staying consistent with list_lru's own memcg migration on
offline.

Thanks,
Qi

[1]. https://lore.kernel.org/all/[email protected]/

Qi Zheng (4):
  fs: fix missed removal of super_fs_objects_eligible()
  mm: memcontrol: make obj_cgroup_memcg() handle NULL objcg
  mm: shmem: move unused huge shrinklist queuing past the truncation
    check
  mm: shmem: make unused huge shrinker memcg aware

 fs/super.c                 |  18 +-
 include/linux/memcontrol.h |  11 +-
 include/linux/shmem_fs.h   |  12 +-
 mm/page_owner.c            |   2 +-
 mm/shmem.c                 | 382 ++++++++++++++++++++++++++++---------
 mm/zswap.c                 |  17 +-
 6 files changed, 315 insertions(+), 127 deletions(-)

-- 
2.54.0