On Tue, 2012-03-06 at 13:56 +0100, Miklos Szeredi wrote:

From: Miklos Szeredi <redacted>

NFSv4 can't do reliable opens in d_revalidate, since it cannot know whether a
mount needs to be followed or not.  It does check d_mountpoint() on the dentry,
which can result in a weird error if the VFS found that the mount does not in
fact need to be followed, e.g.:

  # mount --bind /mnt/nfs /mnt/nfs-clone
  # echo something > /mnt/nfs/tmp/bar
  # echo x > /tmp/file
  # mount --bind /tmp/file /mnt/nfs-clone/tmp/bar
  # cat  /mnt/nfs/tmp/bar
  cat: /mnt/nfs/tmp/bar: Not a directory

Which should, by any sane filesystem, result in "something" being printed.

So instead do the open in f_op->open() and in the unlikely case that the cached
dentry turned out to be invalid, drop the dentry and return ESTALE to let the
VFS retry.

This patch would force a complete new walk of the path in cases where
today we just do a single lookup of the last component. It really
doesn't seem worth taking that penalty just in order to make some insane
bind mount corner cases work.

-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust@netapp.com
www.netapp.com