On Wed, May 08, 2024 at 02:29:05PM +0300, Amir Goldstein wrote: > On Mon, May 6, 2024 at 4:25 PM Chuck Lever <chuck.lever@xxxxxxxxxx> wrote: > > > > It's apparent that a number of distributions and their customers > > remain on long-term stable kernels. We are aware of the scalability > > problems and other bugs in NFSD in kernels between v5.4 and v6.1. > > > > Chuck, > > Are you able to share a partial list of scalability problems that were > fixed by this backport series? > > Specifically, my interest is in the list of improvements to 5.15.y. In broad strokes: - The garbage collection mechanism was rewritten to keep the LRU list short, and to keep sweeps productive. This was the main issue as the LRU used to have hundreds of thousands of files on it and one sweep would take so much CPU it was reported as a soft lockup. - NFSv4 OPEN files are no longer garbage collected so that an NFSv4 CLOSE means local accessors have immediate access to a file. - The filecache hash table is converted to an rhltable so that it can efficiently manage many more open files. - Various fixes prevent writeback of garbage-collected files from bogging down. There are a handful of important bugfixes before this commit, but starting with commit 0369b53886ec ("NFSD: Report filecache LRU size"), have a look at the commits that touch fs/nfsd/filecache.c. -- Chuck Lever