On Tue, Feb 11, 2020 at 12:55:07PM -0500, Johannes Weiner wrote:
> However, this change had to be reverted in 69056ee6a8a3 ("Revert "mm:
> don't reclaim inodes with many attached pages"") because it caused
> severe reclaim performance problems: Inodes that sit on the shrinker
> LRU are attracting reclaim pressure away from the page cache and
> toward the VFS. If we then permanently exempt sizable portions of this
> pool from actually getting reclaimed when looked at, this pressure
> accumulates as deferred shrinker work (a mechanism for *temporarily*
> unreclaimable objects) until it causes mayhem in the VFS cache pools.
>
> In the bug quoted in 69056ee6a8a3 in particular, the excessive
> pressure drove the XFS shrinker into dirty objects, where it caused
> synchronous, IO-bound stalls, even as there was plenty of clean page
> cache that should have been reclaimed instead.
A note on testing: the patch behaves much better on my machine and the
inode shrinker doesn't drop hot page cache anymore, without noticable
downsides so far.
However, I tried to reproduce the xfs case that caused the
69056ee6a8a3 revert and haven't managed to do so yet on 5.5 plus the
reverted patch. I cannot provoke higher inode sync stalls in the xfs
shrinker regardless of shrinker strategy. Maybe something else changed
since 4.19 and it's less of a concern now.
Nonetheless, I'm interested in opinions on the premise of this
patch. And Yafang is working on his memcg-specific fix for this issue,
so I wanted to put this proposal on the table sooner than later.
Thanks
