Re: [PATCH v2 02/31] exportfs: add new EXPORT_OP_STABLE_HANDLES flag

On Mon 19-01-26 23:50:13, Christoph Hellwig wrote:
> On Mon, Jan 19, 2026 at 11:26:19AM -0500, Jeff Layton wrote:
> > +  EXPORT_OP_STABLE_HANDLES - This filesystem provides filehandles that are
> > +    stable across the lifetime of a file. This is a hard requirement for export
> > +    via nfsd. Any filesystem that is eligible to be exported via nfsd must
> > +    indicate this guarantee by setting this flag. Most disk-based filesystems
> > +    can do this naturally. Pseudofilesystems that are for local reporting and
> > +    control (e.g. kernfs, pidfs, nsfs) usually can't support this.
> 
> Suggested rewording, taking some of the ideas from Dave Chinners earlier
> comments into account:
> 
>   EXPORT_OP_STABLE_HANDLES - This filesystem provides filehandles that are
>     stable across the lifetime of a file.  A file in this context is an
>     instantiated inode reachable by one or more file names, or still open after
>     the last name has been unlinked.  Reuses of the same on-disk inode structure
>     are considered new files and must provide different file handles from the
>     previous incarnation.  Most file systems designed to store user data
>     naturally provide this capability.  Pseudofilesystems that are for local
>     reporting and control (e.g. kernfs, pidfs, nsfs) usually can't support this.
> 
>     This flags is a hard requirement for export via nfsd. Any filesystem that
>     is eligible to be exported via nfsd must indicate this guarantee by
>     setting this flag.

I like this. It certainly makes the requirement of stability clearer to me
(with explanations before I couldn't quite see the difference between shmem
and kernfs). I'd note that fat or shmem (which are both exportable)
satisfy this only with reasonably high probability as they use
get_random_u32() for initializing their i_generation but I guess it's as
good as it gets for them.

								Honza
-- 
Jan Kara <[email protected]>
SUSE Labs, CR