Re: [RFC PATCH v5 00/14] mm: mglru: fix swappiness behavior

[Barry Song]

On Sun, Oct 25, 2026 at 12:10 AM Kairui Song <ryncsn@xxxxxxxxx> wrote:
>
> On Fri, Oct 23, 2026 at 6:02 AM Barry Song (Xiaomi) <baohua@xxxxxxxxxx> wrote:
> >
> > Classic LRU respects swappiness, while MGLRU only partially honors it
> > when selecting the scan type, resulting in a much smaller overall
> > impact.
> >
> > For example, when building the kernel in a memcg limited to 1024 MB
> > of memory, with swappiness=200, CLRU barely scans or reclaims file
> > pages, while MGLRU still reclaims a significant number of them.
> > Similarly, with swappiness=1, MGLRU does not reclaim as many file
> > pages as expected for such an extreme swappiness value:
> >
> > pgsteal_file
> > swappiness   CLRU          MGLRU(no patches)
> > 1            5,894,538     756,589
> > 36           799,998       497,282
> > 71           548,331       415,612
> > 106          407,012       383,814
> > 141          253,737       354,249
> > 176          189,003       347,805
> > 200          0             346,078
> >
> > pgsteal_anon
> > swappiness   CLRU          MGLRU(no patches)
> > 1            3,962,404     2,768,654
> > 36           2,830,250     2,976,260
> > 71           3,121,386     3,142,954
> > 106          3,521,720     3,382,383
> > 141          3,504,471     2,970,859
> > 176          3,592,335     3,163,130
> > 200          14,095,617    3,209,132
> >
> > pgscan_file
> > swappiness   CLRU          MGLRU(no patches)
> > 1            10,270,808    886,997
> > 36           1,010,341     569,036
> > 71           656,136       490,972
> > 106          460,326       436,042
> > 141          295,184       403,132
> > 176          235,147       390,358
> > 200          0             398,363
> >
> > pgscan_anon
> > swappiness   CLRU          MGLRU(no patches)
> > 1            19,341,957    5,960,884
> > 36           17,530,774    6,394,951
> > 71           18,624,264    6,673,569
> > 106          20,088,558    6,976,364
> > 141          19,686,813    6,206,580
> > 176          20,936,177    6,344,854
> > 200          45,313,873    6,295,581
> >
> > refault_file
> > swappiness   CLRU          MGLRU(no patches)
> > 1            5,795,115     678,254
> > 36           720,877       419,659
> > 71           469,270       338,518
> > 106          331,584       308,983
> > 141          174,509       277,375
> > 176          113,352       272,390
> > 200          0             269,349
> >
> > refault_anon
> > swappiness   CLRU          MGLRU(no patches)
> > 1            738,618       578,308
> > 36           451,410       665,723
> > 71           514,961       657,870
> > 106          633,161       786,893
> > 141          648,640       662,673
> > 176          668,779       708,969
> > 200          4,816,312     672,514
>
> Hi Barry,
>
> Thanks for pushing this series forward!
>
> One nit pick so far is the test data seems redundant with the one
> below, maybe we can just keep one of them.

make sense to me.

>
> > 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.
>
> Right, the OOM reproducer is demostrated here previously:
> https://lore.kernel.org/all/20260318-mglru-reclaim-v1-0-2c46f9eb0508@xxxxxxxxxxx/
>
> That series is merged at V7:
> https://lore.kernel.org/linux-mm/20260428-mglru-reclaim-v7-0-02fabb92dc43@xxxxxxxxxxx/
>
> The problem is largely eased as described, but it isn't 100% fixed at
> that time. It will still go OOM with higher pressure. So before this
> series, I could still see OOM after hundreds of loops. And problem is
> gone after the fixing patch in this series. Unfortunately it's very
> hard to reproduce and that reproducer is the best I have so far.
>

thanks! I will put this into cover-letter in the new version.

> >
> > Patch 9 from Barry fixes a cold/hot inversion exacerbated by patches
> > 1–8 (which also exists in mainline) and improves reclaim efficiency.
> >
> > Patches 10-13 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 14 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            5,894,538     756,589       7,394,183
> > 36           799,998       497,282       526,176
> > 71           548,331       415,612       433,353
> > 106          407,012       383,814       347,647
> > 141          253,737       354,249       314,437
> > 176          189,003       347,805       261,677
> > 200          0             346,078       188
> >
> > pgsteal_anon
> > swappiness   CLRU          MGLRU         MGLRU+patches
> > 1            3,962,404     2,768,654     5,362,153
> > 36           2,830,250     2,976,260     2,736,749
> > 71           3,121,386     3,142,954     3,246,589
> > 106          3,521,720     3,382,383     3,074,851
> > 141          3,504,471     2,970,859     3,178,409
> > 176          3,592,335     3,163,130     3,650,544
> > 200          14,095,617    3,209,132     14,538,399
> >
> > pgscan_file
> > swappiness   CLRU          MGLRU         MGLRU+patches
> > 1            10,270,808    886,997       16,246,260
> > 36           1,010,341     569,036       614,389
> > 71           656,136       490,972       484,769
> > 106          460,326       436,042       367,969
> > 141          295,184       403,132       328,943
> > 176          235,147       390,358       268,142
> > 200          0             398,363       188
> >
> > pgscan_anon
> > swappiness   CLRU          MGLRU         MGLRU+patches
> > 1            19,341,957    5,960,884     7,646,122
> > 36           17,530,774    6,394,951     5,914,405
> > 71           18,624,264    6,673,569     7,167,762
> > 106          20,088,558    6,976,364     7,821,921
> > 141          19,686,813    6,206,580     8,495,970
> > 176          20,936,177    6,344,854     10,585,375
> > 200          45,313,873    6,295,581     30,313,342
> >
> > refault_file
> > swappiness   CLRU          MGLRU         MGLRU+patches
> > 1            5,795,115     678,254       7,315,064
> > 36           720,877       419,659       448,199
> > 71           469,270       338,518       355,269
> > 106          331,584       308,983       270,736
> > 141          174,509       277,375       237,643
> > 176          113,352       272,390       184,908
> > 200          0             269,349       102
> >
> > refault_anon
> > swappiness   CLRU          MGLRU         MGLRU+patches
> > 1            738,618       578,308       1,383,725
> > 36           451,410       665,723       543,071
> > 71           514,961       657,870       683,874
> > 106          633,161       786,893       669,716
> > 141          648,640       662,673       719,231
> > 176          668,779       708,969       901,918
> > 200          4,816,312     672,514       5,511,787
>
> Right, that's really nice result. Do you have system time or
> performance readings? BTW for the performance reading I think we can
> focus on the 36 - 176 range. If a user asks for extreme bias towards
> anon/file, they strongly favor reclaim priority over throughput, and
> the behavior is arguable at that point.

I think the changes are really minor with the default or
moderate swappiness of around 60, probably just within
the noise of measurement.

With higher swappiness values (>140), I do see an increase
in system time. However, I think this is expected behaviour
as swappiness starts to take effect: we're swapping out and
swapping in more.

real (seconds)
swappiness   MGLRU(no patches)   MGLRU(with patches)
1            110.1               159.2
36           108.8               109.2
71           109.2               109.7
106          109.4               109.2
141          108.4               108.7
176          109.2               109.4
200          108.9               122.8

user (seconds)
swappiness   MGLRU(no patches)   MGLRU(with patches)
1            1519.3              1496.0
36           1531.4              1526.9
71           1530.6              1532.3
106          1533.4              1533.5
141          1536.2              1536.1
176          1530.9              1541.2
200          1536.1              1563.4

sys (seconds)
swappiness   MGLRU(no patches)   MGLRU(with patches)
1            218.3               392.7
36           222.8               216.8
71           221.3               225.3
106          227.0               225.6
141          219.8               231.1
176          221.8               245.1
200          221.1               496.6

As for CLRU, this is also the case:

*** Executing round 1 ***
set swappiness to 36
real 1m50.203s
user 25m23.168s
sys 3m38.826s

*** Executing round 2 ***
set swappiness to 71
real 1m49.654s
user 25m30.618s
sys 3m38.215s

*** Executing round 3 ***
set swappiness to 106
real 1m50.021s
user 25m31.175s
sys 3m51.513s

*** Executing round 4 ***
set swappiness to 141
real 1m50.912s
user 25m37.081s
sys 3m51.816s

*** Executing round 5 ***
set swappiness to 176
real 1m52.159s
user 25m41.676s
sys 3m56.990s
...

Best,
Barry