tree dbc096fb09319a2f04dc5a1c0ae65cdcce0ed135
parent 666b3d803a511fbc9bc5e5ea8ce66010cf03ea13
author Trond Myklebust <Trond.Myklebust@netapp.com> 1361555623 -0500
committer Trond Myklebust <Trond.Myklebust@netapp.com> 1361562934 -0500

NFS: Don't allow NFS silly-renamed files to be deleted, no signal

Commit 73ca100 broke the code that prevents the client from deleting
a silly renamed dentry.  This affected "delete on last close"
semantics as after that commit, nothing prevented removal of
silly-renamed files.  As a result, a process holding a file open
could easily get an ESTALE on the file in a directory where some
other process issued 'rm -rf some_dir_containing_the_file' twice.
Before the commit, any attempt at unlinking silly renamed files would
fail inside may_delete() with -EBUSY because of the
DCACHE_NFSFS_RENAMED flag.  The following testcase demonstrates
the problem:
  tail -f /nfsmnt/dir/file &
  rm -rf /nfsmnt/dir
  rm -rf /nfsmnt/dir
  # second removal does not fail, 'tail' process receives ESTALE

The problem with the above commit is that it unhashes the old and
new dentries from the lookup path, even in the normal case when
a signal is not encountered and it would have been safe to call
d_move.  Unfortunately the old dentry has the special
DCACHE_NFSFS_RENAMED flag set on it.  Unhashing has the
side-effect that future lookups call d_alloc(), allocating a new
dentry without the special flag for any silly-renamed files.  As a
result, subsequent calls to unlink silly renamed files do not fail
but allow the removal to go through.  This will result in ESTALE
errors for any other process doing operations on the file.

To fix this, go back to using d_move on success.
For the signal case, it's unclear what we may safely do beyond d_drop.

Reported-by: Dave Wysochanski <dwysocha@redhat.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Acked-by: Jeff Layton <jlayton@redhat.com>
Cc: stable@vger.kernel.org