[Unionfs] mode=nfsro (Read only NFS) in 2.0

[Erez Zadok]

In message <469BB369.4040506 at hellosam.net>, Sam Wong writes:
> Hi,
> 
> I would like to know if mode=nfsro is gone or even in a "won't fix"
> status? It was working in Union 1.x but unfortunately I have to stick
> with a recent kernel.
> 
> As far as I can see, the NFSRO defines is still left in the inodes.c, I
> would suspect that it would be easy to add the support back into the
> code. Or am I totally wrong about that?
> 
> Many Thanks,
> Sam

Sam, I tested the latest Unionfs code with a Linux 2.6.22 NFSv3 client and
server.  I tried various options mixing readonly and readwrite exports and
mounts.  As of now, I believe that the older mode=nfsro option is no longer
necessary, as the NFS client/server code nowadays do the right thing.  In
other words, I don't think you need it at all and everything "should" work
right.

Explanation: for Unionfs to perform a copyup, it needs to know that a branch
is readonly.  That's easy when the branch itself is mounted readonly, or the
underlying f/s is mounted readonly.  However, if the branch is mounted
readwrite, but comes from an NFS server which exports the f/s as readonly,
we used to have a problem.  The problem was that the NFS client/server code,
in the case of a readonly-exported volume, was returning -EACCES instead of
-EROFS.  An EACCES error should not trigger a copyup, as that would be a
security violation.  So we had to create a special mount option which
translates EACCES on an NFS branch, to an EROFS.  That option was a hack at
best; the real answer was to get the NFS code to behave more consistently,
and it looks like it does so now (Jeff and I had a few close chats with the
NFS maintainers a few months ago -- looks like it worked :-)

You're right in that there's still some old NFSRO code left over in
unionfs2, and we might just remove it.

That said, I'd like to be sure before I remove code that may be necessary.
After all, I only tried this with a recent Linux NFS server and client
(2.6.22).  But what about other NFS servers (Solaris? BSD? etc.)

I'd like to ask those of you on this list: do you know of situations with
the latest unionfs2 code, where you *do* need an "nfsro" mount option?  Or
does unionfs2 work for you ok even with a readonly-exported NFS volume?

If it doesn't work for you, and you find that you do need nfsro support,
then please give me enough details so I can reproduce this problem on my end
and see how to best resolve it.  Otherwise I'd like to remove the remains of
this hacky code.

Thanks,
Erez.