On Sun, Jan 24, 2021 at 05:18:54PM -0500, J. Bruce Fields wrote:

On Sat, Jan 23, 2021 at 02:09:58PM +0100, Christian Brauner wrote:

quoted

On Fri, Jan 22, 2021 at 05:26:32PM -0500, J. Bruce Fields wrote:

quoted

If I NFS-exported an idmapped mount, I think I'd expect idmapped clients
to see the mapped IDs.

Looks like that means taking the user namespace from the struct
svc_export everwhere, for example:

On Thu, Jan 21, 2021 at 02:19:24PM +0100, Christian Brauner wrote:

quoted

index 66f2ef67792a..8d90796e236a 100644

--- a/fs/nfsd/nfsfh.c
+++ b/fs/nfsd/nfsfh.c

@@ -40,7 +40,8 @@ static int nfsd_acceptable(void *expv, struct dentry *dentry)
 		/* make sure parents give x permission to user */
 		int err;
 		parent = dget_parent(tdentry);
-		err = inode_permission(d_inode(parent), MAY_EXEC);
+		err = inode_permission(&init_user_ns,
+				       d_inode(parent), MAY_EXEC);

		err = inode_permission(exp->ex_path.mnt->mnt_userns,
				      d_inode(parent, MAY_EXEC);

Hey Bruce, thanks! Imho, the clean approach for now is to not export
idmapped mounts until we have ported that part of nfs similar to what we
do for stacking filesystems for now. I've tested and taken this patch
into my tree:

Oh good, thanks.  My real fear was that we'd fix this up later and leave
users in a situation where the server exposes different IDs depending on
kernel version, which would be a mess.  Looks like this should avoid
that.

As for making idmapped mounts actually work with nfsd--are you planning
to do that, or do you need me to?  I hope the patch is straightforward;

I'm happy to do it or help and there's other people I know who are also
interested in that and would likely be happy to do the work too.

I'm more worried testing it.

This whole series has a large xfstest patch associated with it that
tests regular vfs behavior and vfs behavior with idmapped mounts. Iirc,
xfstests also has infrastructure to test nfs. So I'd expect we expand
the idmapped mounts testsuite to test nfs behavior as well.
So far it has proven pretty helpful and has already unconvered an
unrelated setgid-inheritance xfs bug that Christoph fixed a short time
ago.