Re: [RFC][PATCH] nfsd: vfs_mkdir() might succeed leaving dentry negative unhashed

["J. Bruce Fields" <bfields@fieldses.org>]

On Mon, May 14, 2018 at 04:45:51PM +0100, Al Viro wrote:
> On Mon, May 14, 2018 at 11:32:16AM -0400, J. Bruce Fields wrote:
> > On Fri, May 11, 2018 at 10:13:39PM +0100, Al Viro wrote:
> > > [-stable fodder; as it is, one can e.g. add
> > > /mnt/cgroup     localhost(rw,no_root_squash,fsid=4242)
> > > to /etc/exports,
> > > mount -t cgroup none /mnt/cgroup
> > > mkdir /tmp/a
> > > mount -t nfs localhost://mnt/cgroup /tmp/a
> > > mkdir /tmp/a/foo
> > 
> > How is the cgroup filesystem exportable?  That sounds like a bug in
> > itself.  We don't want people using NFS as some kind of weird remote
> > configuration protocol.
> 
> You can't have open-by-fhandle without exportability.  And it's not
> the only fs like that.

We could separate the two--add a flag to export_operations, if
necessary.  I haven't formulated a strong argument, but exporting those
filesystems makes me *really* uncomfortable.

Poking around....  Looks like this was added by aa8188253474 "kernfs:
add exportfs operations", and they really do claim a use case for lookup
by filehandle.

> > > and have knfsd oops; the patch below deals with that.
> > > 
> > > Questions:
> > > 	1) is fh_update() use below legitimate, or should we
> > > do fh_put/fh_compose instead?
> > 
> > fh_update looks OK to me, but do we need it here?  There's already a
> > 
> > 	if (!err)
> > 		err = fh_update(reshp);
> > 
> > at the end of nfsd_create_locked.
> 
> Might be too late for that, though - the trouble hits when we hit
> nfsd_create_setattr().

Oh, got it.  Could move the bottom fh_update to just above the
nfsd_create_setattr(), though?

> > > 	2) is nfserr_serverfail valid for situation when
> > > directory created by such vfs_mkdir() manages to disappear
> > > under us immediately afterwards?  Or should we return nfserr_stale
> > > instead?
> > 
> > We just got a successful result on the create and the parent's still
> > locked, so if somebody hits this I think we want them reporting a bug,
> > not wasting time trying to find a mistake in their own logic.
> 
> No.  Suppose it's NFS-over-NFS (and yes, I agree that it's a bad idea;
> somebody *has* done that).  Lookup after successful mkdir can legitimately
> fail if it's been removed server-side.
> 
> And we *will* need to allow nfs_mkdir() to return that way in some cases
> (== success with dentry passed to it left unhashed negative), I'm afraid ;-/

Thanks, makes sense.

--b.