On Wed, 22 May 2024 at 01:12, kernel test robot <oliver.sang@xxxxxxxxx> wrote:
>
> kernel test robot noticed a 6.7% improvement of stress-ng.touch.ops_per_sec on:
>
> commit: 3681ce364442ce2ec7c7fbe90ad0aca651db475b ("[PATCH] vfs: Delete the associated dentry when deleting a file")
Ok, since everything else is at least tentatively in the noise, and
the only hard numbers we have are the ones from Yafang's Elasticsearch
load and this - both of which say that this is a good patch - I
decided to just apply this ASAP just to get more testing.
I just wanted to note very explicitly that this is very much
tentative: this will be reverted very aggressively if somebody reports
some other real-world load performance issues, and we'll have to look
at other solutions. But I just don't think we'll get much more actual
testing of this without just saying "let's try it".
Also, I ended up removing the part of the patch that stopped clearing
the DCACHE_CANT_MOUNT bit. I think it's right, but it's really
unrelated to the actual problem at hand, and there are other cleanups
- like the unnecessary dget/dput pair - in this path that could also
be looked at.
Anyway, let's see if somebody notices any issues with this. And I
think we should look at the "shrink dentries" case anyway for other
reasons, since it's easy to create a ton of negative dentries with
just lots of lookups (rather than lots of unlinking of existing
files).
Of course, if you do billions of lookups of different files that do
not exist in the same directory, I suspect you just have yourself to
blame, so the "lots of negative lookups" load doesn't sound
particularly realistic.
TLDR; I applied it for testing because we're in the merge window and
there's no reason not to try.
Linus
