[PATCH v9 00/16] mm: mglru: fix swappiness behavior

From: Barry Song (Xiaomi)

Date: Sat Nov 14 2026 - 15:08:22 EST


Classic LRU respects swappiness, while MGLRU only weakly
honors it when selecting the scan type. The fallback logic
can easily override the bias established by
get_type_to_scan(). As a result, swappiness has a much
smaller overall impact when changing from 1 to 200, as
shown by the data presented later in this cover letter.

In this patchset, patches 1-8 from Kairui introduce a model that uses
refaults to determine nr_to_scan[] for both file and anon in MGLRU.
For each reclaim operation, we calculate how many file and anon
folios to scan, similar to CLRU, rather than relying on
get_type_to_scan() and fallback mechanisms to decide what to
reclaim, which have proven rather unreliable in honoring swappiness.
This model also allows scanning and reclaiming from the two youngest
generations, including the youngest generation itself. This can help
resolve OOM situations[1] and reduce the need for aging.

As Kairui mentioned, the OOM reproducer was demonstrated
previously [2], and that series was merged in v7 [3].
The problem was largely alleviated as described, but not
completely fixed. Under higher memory pressure, the system
could still hit OOM. Before this series, Kairui could still
reproduce OOM after hundreds of iterations. With the fixing
patch in this series applied, the problem no longer occurs.

Patch 9 from Barry fixes a cold/hot inversion exacerbated by patches
1–8 (which also exists in mainline) and improves reclaim efficiency.

Patches 10-15 from Gaofei and Barry optimize aging based on folio
distribution imbalance and improve reclaim behavior at extreme
swappiness values such as 1 and 200.

Patch 16 from Bo Zhang addresses the GFP_NOIO issue in MGLRU.

After the patchset, the results track CLRU across the full swappiness
range (see table below), with more file scanning at low swappiness
and more anon scanning at high swappiness.

pgsteal_file

swappiness CLRU MGLRU MGLRU+patches
1 6,620,680 745,968 6,197,601
36 807,807 508,974 564,256
71 553,144 417,700 420,008
106 402,880 402,474 363,635
141 240,293 337,797 312,624
176 187,682 330,045 238,722
200 0 343,949 0


pgsteal_anon

swappiness CLRU MGLRU MGLRU+patches
1 4,436,199 2,671,934 5,011,932
36 2,858,117 3,032,747 3,039,483
71 3,042,392 2,827,735 3,132,431
106 3,292,622 3,184,314 3,238,658
141 3,612,869 2,999,000 3,306,901
176 3,827,154 3,134,127 3,714,411
200 13,825,283 3,269,753 14,481,445


pgscan_file

swappiness CLRU MGLRU MGLRU+patches
1 11,853,625 877,311 13,108,154
36 1,016,826 582,783 657,973
71 662,823 474,772 457,686
106 496,374 451,285 390,056
141 277,109 382,604 329,819
176 206,261 372,059 246,141
200 0 393,608 0


pgscan_anon

swappiness CLRU MGLRU MGLRU+patches
1 20,804,386 5,847,747 6,936,021
36 17,830,330 6,740,075 5,975,744
71 18,357,293 6,126,194 6,797,243
106 19,938,920 6,431,355 7,602,990
141 19,609,579 5,997,762 8,540,965
176 21,612,749 6,071,032 10,532,510
200 44,776,474 6,544,125 30,296,434


refault_file

swappiness CLRU MGLRU MGLRU+patches
1 6,523,070 667,302 6,118,731
36 722,110 431,585 486,604
71 477,062 340,241 342,746
106 327,610 324,747 286,465
141 161,827 265,959 233,860
176 115,551 258,515 171,174
200 0 268,049 0


refault_anon

swappiness CLRU MGLRU MGLRU+patches
1 867,728 552,829 1,368,563
36 419,635 657,745 636,567
71 536,299 611,345 689,528
106 580,742 693,689 707,227
141 706,151 639,850 753,779
176 706,499 685,086 934,553
200 4,693,850 733,273 5,491,234

With a normal swappiness setting (without extreme bias), I
see almost no noticeable performance difference between
the two versions.

For example, with `swappiness=60`, the average results
over 5 rounds are:

metric W/o patch W/ patch
real 1m49.273s 1m48.836s
user 25m33.186s 25m33.368s
sys 3m41.851s 3m38.388s

[1] https://github.com/ryncsn/emm-test-project/tree/master/file-anon-mix-pressure
[2] https://lore.kernel.org/all/20260318-mglru-reclaim-v1-0-2c46f9eb0508@xxxxxxxxxxx/
[3] https://lore.kernel.org/linux-mm/20260428-mglru-reclaim-v7-0-02fabb92dc43@xxxxxxxxxxx/

v9:
* reduce one iteration for lruvec_populated_min_seq();
* let lruvec_is_reclaimable() count all zones, for min_ttl;
* refine lru_gen_imbalanced(), AGING_PRIORITY macro, roundup/down swappiness;
per Kairui;
* let evict_folio_lists() return "drained" per Kairui, thanks!
* one bounded retry for scanned=0 and reduce concurrent aging
for no-populatable(drained) aging;
v8 link:
https://lore.kernel.org/linux-mm/20261113185541.70203-1-baohua@xxxxxxxxxx/

v8:
* Fixed Gaofei Lyu's name;
* for debugfs proactive reclaim, use DEF_PRIORITY;
* refine no-populatable aging (limit frequency, allow an aging
after eviction returns 0 scanned)
* refine "restrict eviction accounting to eligible zones" for
potential premature OOM;
* Collect Baolin's reviewed-by tag, thanks!
v7 link:
https://lore.kernel.org/linux-mm/20261111232603.9934-1-baohua@xxxxxxxxxx/

v7:
* fix a build error of bisect and address some other sashiko minor comments;
* add no-populatable aging;
v6 link:
https://lore.kernel.org/linux-mm/20261110202025.25736-1-baohua@xxxxxxxxxx/

v6:
* rebase on top of mm-new(a lot of conflicts due to recent active mglru
developments);
* only count eligible zones below sc->reclaim_idx for
lruvec_evictable_size() and lru_gen_imbalanced();
* GFP_NOIO further fixing by Bo Zhang, exclude unreclaimable type
in lruvec_evictable_size().
* Add Kunwu's reviewed-by, thanks!
* RFC v5 link:
https://lore.kernel.org/linux-mm/20261022220203.91911-1-baohua@xxxxxxxxxx/

RFC v5:
* Remove the aging optimizations, which have been merged - thanks for Baolin,
Baoqun, Lian Wang, Kairui, Kunwu, Ridong reviewing on this;
* Add Kairui's model for determining nr_to_scan[] for both types.
* Fix cold/hot inversion for rejected folios - Barry.
* Improve and clean up imbalanced aging, and optimize extreme
swappiness values of 1 and 200 - Gaofei, Barry.
* Address the GFP_NOIO issue in MGLRU - Bo Zhang.
* RFC v4 link:
https://lore.kernel.org/linux-mm/20260812121658.69965-1-baohua@xxxxxxxxxx/

RFC v4:
* Clean up isolate_folios() based on Baolin's feedback. Thanks!
* Further improve isolate_folios() and scan_folios(), including
handling of exhausted reclaim types.
* Add fixes and performance improvements to inc_min_seq(),
and reduce lock contention.
* Improve the handling of imbalanced aging and remaining
generation-based aging.
* RFC v3 link:
https://lore.kernel.org/linux-mm/20260731083843.37811-1-baohua@xxxxxxxxxx/

RFC v3:
- Added Ridong's patch of improving isolate_folios() readability;
- for lru_gen_imbalanced(), apply some suggestions from Zicheng
and Kairui, considering lruvec size and use macro;
- drop "mm: mglru: only fall back when reclaim is running at high
priority"
- mitigate aging for balanced swappiess value(near middle 100)
as swappiness bias is less a concern for them; but slightly
increased aging overhead might not be good.
* RFC v2 link:
https://lore.kernel.org/linux-mm/20260726122123.7614-1-baohua@xxxxxxxxxx/

RFC v2:
- Quickly address a few issues raised in the Sashiko comments so
reviewers can ignore v1 and review a cleaner version instead.
https://sashiko.dev/#/patchset/20260726012946.18684-1-baohua@xxxxxxxxxx
Thanks, Sashiko!

Barry Song (Xiaomi) (5):
mm/mglru: Avoid placing rejected folios below populated gen
mm/mglru: Run aging when one type is drained
mm/mglru: Restrict eviction accounting to eligible zones
mm/mglru: Reduce LRU lock contention by shrinking folio batches in
inc_min_seq
mm/mglru: Make swappiness=1 file reclaim more aggressive

Bo Zhang (1):
mm/mglru: Avoid scan anon while swapcache is low for GFP_NOIO
allocation

Gaofei Lyu (2):
mm/mglru: Run aging when pages are rare outside the youngest gen
mm/mglru: Avoid prolonged anon reclaim with swappiness=200

Kairui Song (8):
mm/mglru: report evictable sizes per type
mm/mglru: extract lruvec_populated_min_seq() helper
mm/mglru: reclaim from the oldest populated gen
mm/mglru: lift the forced protection of reclaimer
mm/mglru: skip zones above reclaim_idx instead of rotating them
mm/mglru: return for unavailable anon reclaim earlier
mm/mglru: balance anon and file scan budget by refault rate
mm/mglru: lift youngest gen protection for higher priority

include/linux/mm_inline.h | 22 +-
mm/trace_vmscan.h | 4 +-
mm/vmscan.c | 514 +++++++++++++++++++++++++-------------
mm/vmscan.h | 7 +-
4 files changed, 364 insertions(+), 183 deletions(-)

--
2.39.3 (Apple Git-146)