[RFC PATCH 0/6] mm/mglru: skip empty PUD subtrees during aging with PUD-level Bloom filter

[Baoquan He <baoquan.he@linux.dev>]

Problem
=======

MGLRU's aging walks every present PUD of every mm in the mm_list.  When
an mm has no pages on a given NUMA node, all 512 PMDs in each PUD fail
the existing PMD-level Bloom filter test, yet the walker still descends
every PUD and iterates every PMD — pure waste.

These cross-node empty walks are structural: lru_gen_use_mm() sets
mm->lru_gen.bitmap to -1 (all nodes) at each context switch, so every
node's kswapd independently walks the same mm.  On a 2-node KVM guest,
~80% of all aging walks are empty; the worst-case leaf-traversal waste
with a fully resident remote mm is ~73%.

Approach
========

Add a PUD-level Bloom filter (one level above the existing PMD filter).
The walker now:

  1. Checks the PUD filter before descending into a PUD subtree.
  2. If the filter says the PUD had no young entries last generation,
     skips the entire 1GB region — avoiding 512 PMD lookups.
  3. walk_pmd_range() reports upward whether it found any young leaf
     entries; walk_pud_range() records that in the double-buffered
     PUD filter.
  4. The rmap feedback path (lru_gen_look_around()) marks hot PUDs in
     the filter, so newly hot or migrated-in regions are re-examined
     promptly rather than suppressed indefinitely.
  5. force_scan walks bypass the PUD test, so manual aging and newly
     added mm's always populate the filter.

The double-buffered filter flips each aging generation, and the eviction
feedback keeps hot regions marked, so correctness holds naturally without
tracking per-page residency or requiring changes to page fault / rmap
hot paths.

The testing results are from a 2-node KVM guest.  Real-hardware results on
a  ≥4-socket system, would be valuable. I will update the statistics
here if have.

Baoquan He (6):
  mm/mglru: add PUD-level Bloom filter state
  mm/mglru: refactor Bloom filter helpers for two filter levels
  mm/mglru: skip empty PUD subtrees during aging
  mm/mglru: report hot PUDs from the rmap feedback path
  mm/mglru: add MM_WALK_EMPTY stats and tracepoint for cross-node
    measurement
  mm/mglru: count PUD subtrees skipped by the PUD-level filter

 include/linux/mmzone.h        |  15 +++-
 include/trace/events/vmscan.h |  30 ++++++++
 mm/vmscan.c                   | 126 +++++++++++++++++++++++++++++-----
 3 files changed, 153 insertions(+), 18 deletions(-)

--
2.54.0