Re: [PATCH v8] fuse: add more control over cache invalidation behaviour

[Miklos Szeredi]

On Mon, 10 Mar 2025 at 21:11, Bernd Schubert <bernd@xxxxxxxxxxx> wrote:

> Can't that be done in fuse-server? Maybe we should improve
> notifications to allow a batch of invalidations?
>
> I'm a bit thinking about
> https://github.com/libfuse/libfuse/issues/1131
>
> I.e. userspace got out of FDs and my guess is it happens
> because of dentry/inode cache in the kernel. Here userspace
> could basically need to create its own LRU and then send
> invalidations. It also could be done in kernel,
> but kernel does not know amount of max open userspace FDs.
> We could add it into init-reply, but wouldn't be better
> to keep what we can in userspace?

Two different things:

1) trimming the cache: this is what you are taking about above.  I
don't think it's possible to move the LRU to userspace since it
doesn't see cache accesses and also does not have information about
some references (e.g. cwd). This can be solved by adding a
notification (FUSE_NOTIFY_TRIM) that tell the kernel to evict  N
*unused* dentries (valid or invalid).

2) cleaning up of invalid dentries.  Dentries can become invalid by
explicit invalidation or by expiring the timeout.  The latter is a bit
of a challenge to clean up, as we don't want to start a timer for each
dentry.   This is what I was suggesting instead of an explicit
shrink_dcache_sb().

Thanks,
Miklos