On Mon, Dec 03, 2018 at 11:12:59AM +0100, Ondrej Mosnacek wrote:

I think I figured out what's the problem. NFS still creates the
submount via the old vfs_submount() call, which calls
vfs_kern_mount(), which creates an fs_context with
FS_CONTEXT_FOR_USER_MOUNT because FS_CONTEXT_FOR_SUBMOUNT needs the
mountpoint dentry reference and there is currently no way to pass that
to vfs_kern_mount(). This is further complicated by the fact that
vfs_submount() accepts only a const reference to the mountpoint, while
vfs_new_fs_context() expects a non-const one...

I think all users of the old vfs_submount call should be converted to
the new API before the VFS changes are merged into mainline, otherwise
they will break the SELinux submount fix. We could work around it in
the SELinux hook by checking the fc->sb_flags[_mask] for SB_SUBMOUNT,
but I guess that would be a hack.

Could you take a look at vfs.git#Q28?  There's still a massive reshuffling
going on, so there will be more branches; this one is the latest at the
moment.

I really hate the situation around sb_clone_mnt_opts/sb_set_mnt_opts and
I'm none too fond of the way fs_context_validate is done, so there will
be quite a bit of LSM tweaking.  If we are doing that, let's do it
right...