On Wed, 2023-09-27 at 12:10 +1000, NeilBrown wrote:
>
> hi Trond,
> I'm trying to understand
>
> Commit 76c21e3f70a8 ("mountd: Check the stat() return values in
> match_fsid()")
>
> in nfs-utils.
>
> The effect of this patch is that if a 'stat' of any path in
> /etc/exports or any mountpoint below any path marked crossmnt fails
> with an error other than one of a small set, then the fsid to path
> lookup aborts without reporting anything to the kernel, so the
> kernel
> doesn't reply to the client and the mount attempt blocks
> indefinitely.
>
> I have seen this happen when "/" is exported crossmnt, and when a
> stat
> of /run/user/1000/doc returns EACCES. This is a "fuse" mount for
> user
> 1000, and presumably it rejects any access from any other user.
>
> Could you please help me understand what this patch was trying to
> achieve? What sorts of errors were you expecting this to catch?
> Would it make sense to silently ignore the stat failure for paths
> that
> were found when scanning the mount table, and only take the more
> drastic action for paths explicitly listed in /etc/exports ??
>
>
The point is that if the filesystem is re-exported NFS, and you mount
as 'softerr' in order to avoid hanging your knfsd server on ephemeral
errors, then you have to be more careful about which errors are fatal,
and hence need to be propagated to the kernel export/path cache, and
which ones are just temporary due to a network partition or server
reboot.
Hence the creation of path_lookup_error() which differentiates between
the two cases.
--
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@xxxxxxxxxxxxxxx
