]> www.infradead.org Git - users/dwmw2/linux.git/commit
mm: vmscan: fix kswapd exit condition in defrag_mode
authorJohannes Weiner <hannes@cmpxchg.org>
Wed, 16 Apr 2025 13:45:40 +0000 (09:45 -0400)
committerAndrew Morton <akpm@linux-foundation.org>
Fri, 18 Apr 2025 03:10:09 +0000 (20:10 -0700)
commita1f0220f3319057b364d871659ef7c10ab78f795
tree9b8296b3894923339058fb937a538dd37f1274e2
parent38448181459e24257b40d5258afdbaa3565e8cfc
mm: vmscan: fix kswapd exit condition in defrag_mode

Vlastimil points out an issue with kswapd in defrag_mode not waking up
kcompactd reliably.

Background: When kswapd is woken for any higher-order request, it
initially checks those high-order watermarks to decide if work is
necesary.  However, it cannot (efficiently) meet the contiguity goal of
such a request by itself.  So once it has reclaimed a compaction gap, it
adjusts the request down to check for free order-0 pages, then wakes
kcompactd to coalesce them into larger blocks.

In defrag_mode, the initial watermark check needs to be analogously
against free pageblocks.  However, once kswapd drops the high-order to
hand off contiguity work, it also needs to fall back to base page
watermarks - otherwise it'll keep reclaiming until blocks are freed.

While it appears kcompactd is woken up frequently enough to do most of the
compaction work, kswapd ends up overreclaiming by quite a bit:

                                                     DEFRAGMODE     DEFRAGMODE-thispatch
Hugealloc Time mean                       79381.34 (    +0.00%)    88126.12 (   +11.02%)
Hugealloc Time stddev                     85852.16 (    +0.00%)   135366.75 (   +57.67%)
Kbuild Real time                            249.35 (    +0.00%)      226.71 (    -9.04%)
Kbuild User time                           1249.16 (    +0.00%)     1249.37 (    +0.02%)
Kbuild System time                          171.76 (    +0.00%)      166.93 (    -2.79%)
THP fault alloc                           51666.87 (    +0.00%)    52685.60 (    +1.97%)
THP fault fallback                        16970.00 (    +0.00%)    15951.87 (    -6.00%)
Direct compact fail                         166.53 (    +0.00%)      178.93 (    +7.40%)
Direct compact success                       17.13 (    +0.00%)        4.13 (   -71.69%)
Compact daemon scanned migrate          3095413.33 (    +0.00%)  9231239.53 (  +198.22%)
Compact daemon scanned free             2155966.53 (    +0.00%)  7053692.87 (  +227.17%)
Compact direct scanned migrate           265642.47 (    +0.00%)    68388.33 (   -74.26%)
Compact direct scanned free              130252.60 (    +0.00%)    55634.87 (   -57.29%)
Compact total migrate scanned           3361055.80 (    +0.00%)  9299627.87 (  +176.69%)
Compact total free scanned              2286219.13 (    +0.00%)  7109327.73 (  +210.96%)
Alloc stall                                1890.80 (    +0.00%)     6297.60 (  +232.94%)
Pages kswapd scanned                    9043558.80 (    +0.00%)  5952576.73 (   -34.18%)
Pages kswapd reclaimed                  1891708.67 (    +0.00%)  1030645.00 (   -45.52%)
Pages direct scanned                    1017090.60 (    +0.00%)  2688047.60 (  +164.29%)
Pages direct reclaimed                    92682.60 (    +0.00%)   309770.53 (  +234.22%)
Pages total scanned                    10060649.40 (    +0.00%)  8640624.33 (   -14.11%)
Pages total reclaimed                   1984391.27 (    +0.00%)  1340415.53 (   -32.45%)
Swap out                                 884585.73 (    +0.00%)   417781.93 (   -52.77%)
Swap in                                  287106.27 (    +0.00%)    95589.73 (   -66.71%)
File refaults                            551697.60 (    +0.00%)   426474.80 (   -22.70%)

Link: https://lkml.kernel.org/r/20250416135142.778933-3-hannes@cmpxchg.org
Fixes: a211c6550efc ("mm: page_alloc: defrag_mode kswapd/kcompactd watermarks")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reported-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Brendan Jackman <jackmanb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
mm/vmscan.c