[SRU][N][O][PATCH 0/1] MGLRU: kswapd uses 100% CPU when MGLRU is enabled and under memory pressure

[Matthew Ruffell]

BugLink: https://bugs.launchpad.net/bugs/2087886

[Impact]

On systems with MGLRU enabled, which it is by default, if the system is under
memory pressure, and some pages are then allocated, such that it wakes up kswapd
to attempt page reclaim, but the system has enough memory that kswapd doesn't
get OOM killed, and then there are no pages to actually reclaim, kswapd can spin
at 100% endlessly, causing severe performance issues.

What's happening is that lru_gen_shrink_node() unconditionally clears 
kswapd_failures, which can prevent kswapd from sleeping and cause 100% kswapd 
cpu usage even when kswapd repeatedly fails to make progress in reclaim.

A workaround is to disable MGLRU, and the issue does not occur.

[Fix]

The fix is to only clear kswap_failures in lru_gen_shrink_node() if reclaim 
makes some progress, similar to shrink_node().

This was fixed in 6.12-rc4 by the commit:

commit b130ba4a6259f6b64d8af15e9e7ab1e912bcb7ad
Author: Wei Xu <weixugc at google.com>
Date:   Mon Oct 14 22:12:11 2024 +0000
Subject: mm/mglru: only clear kswapd_failures if reclaimable
Link: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=b130ba4a6259f6b64d8af15e9e7ab1e912bcb7ad

Both Noble and Oracular need this fix.

[Testcase]

The systems with this issue are DPDK compute nodes running OpenStack Yoga on
Jammy. If you leave them for several days, they will hit this issue and their
kswap processes will go to 100% and never drop.

    PID USER      PR  NI    VIRT    RES    SHR S  %CPU  %MEM     TIME+ COMMAND
    846 root      20   0       0      0      0 R 100.0   0.0   1915:16 kswapd0
    846 root      20   0       0      0      0 R 100.0   0.0   1915:16 kswapd0
    846 root      20   0       0      0      0 R 100.0   0.0   1915:17 kswapd0
    846 root      20   0       0      0      0 R  99.0   0.0   1915:18 kswapd0

If you disable MGLRU however, they drop to 0% and sleep immediately.

There is a test kernel available in the following ppa:

https://launchpad.net/~mruffell/+archive/ubuntu/sf390959-test

If you install the test kernel, the system should remain stable, and kswapd will
not go to 100% cpu after some time.

[Where problems could occur]

We are adding a small requirement to clearing kswapd_failures, that we actually
reclaim some memory. If we do actually reclaim some memory, the behaviour is the
same as what we have now, we clear kswapd_failures and continue. 

If we don't manage to reclaim any memory, we leave kswapd_failures as is, and
let kswapd sleep, to try again at a later time.

If a regression were to occur, it would affect MGLRU users, which is the default.
A regression would look like issues with kswapd memory reclaim, or lead to high
cpu usage with kswapd.

Wei Xu (1):
  mm/mglru: only clear kswapd_failures if reclaimable

 mm/vmscan.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

-- 
2.45.2