[Devel] [PATCH RHEL v2] mm: Reduce access frequency to shrinker_rwsem during shrink_slab

[Andrey Ryabinin]

On 8/20/20 5:51 PM, Valeriy Vdovin wrote:
> Bug https://jira.sw.ru/browse/PSBM-99181 has introduced a problem: when
> the kernel has opened NFS delegations and NFS server is not accessible
> at the time when NFS shrinker is called, the whole shrinker list
> execution gets stuck until NFS server is back. Being a problem in itself
> it also introduces bigger problem - during that hang, the shrinker_rwsem
> also gets locked, consequently no new mounts can be done at that time
> because new superblock tries to register it's own shrinker and also gets
> stuck at aquiring shrinker_rwsem.
> 
> Commit 9e9e35d050955648449498827deb2d43be0564e1 is a workaround for that
> problem. It is known that during signle shrinker execution we do not
> actually need to hold shrinker_rwsem so we release and reacqiure the
> rwsem for each shrinker in the list.
> 
> Because of this workaround shrink_slab function now experiences a major
> slowdown, because shrinker_rwsem gets accessed for each shrinker in the
> list twice. On an idle fresh-booted system shrinker_list could be
> iterated up to 1600 times a second, although originally the problem was
> local to only one NFS shrinker.
> 
> This patch fixes commit 9e9e35d050955648449498827deb2d43be0564e1 in a
> way that before calling for up_read for shrinker_rwsem, we check that
> this is really an NFS shrinker by checking NFS magic in superblock, if
> it is accessible from shrinker.
> 
> https://jira.sw.ru/browse/PSBM-99181
> 
> Co-authored-by: Andrey Ryabinin <aryabinin at virtuozzo.com>
> Signed-off-by: Valeriy Vdovin <valeriy.vdovin at virtuozzo.com>
> 
> Changes:
>   v2: Added missing 'rwsem_is_contented' check
> ---

Reviewed-by: Andrey Ryabinin <aryabinin at virtuozzo.com>