Re: [fuse-devel] dentry expiry instead of invalidation

[Miklos S. <mi...@sz...>]

On Sat, 3 Sept 2022 at 01:00, Jakob Blomer <jb...@ce...> wrote:
>
> Dear all,
>
> For the development of cvmfs [1], we are wondering if it is possible to
> let a fuse module expire a certain dentry, so that the next access
> triggers a LOOKUP callback.
>
> The background is this: cvmfs is a read-only network file system.  It is
> not immutable -- every now and then a new file system snapshot becomes
> available remotely and is then served by clients.
>
> We recently started looking into symlink kernel caching.  When a symlink
> is updated to a longer one, the client can temporarily serve a truncated
> version of the new, long symlink.  This is due to the kernel cache still
> using the old entry size.
>
> We know that the following options prevent this behavior:
>
>    - wait for the kernel cache timeout; this is not ideal because it
> prevents clients from swiftly following remote changes (or, with short
> timeouts, makes inefficient use of the kernel cache)
>
>    - Invalidate the dentry with fuse_lowlevel_notify_inval_entry()
> before applying the update.
>
> Invalidating the dentry, however, has at least one undesirable
> side-effect: a mountpoint on the symlink gets irrecoverably destroyed.
>
> We would ideally like to be able to immediately "expire" the dentry, so
> that it remains valid but the next access should trigger a LOOKUP callback.

Makes sense.

Attaching untested patches.

Thanks,
Miklos