rename(): avoid a deadlock in the case of parents having no common ancestor

... and fix the directory locking documentation and proof of correctness.
Holding ->s_vfs_rename_mutex *almost* prevents ->d_parent changes; the
case where we really don't want it is splicing the root of disconnected
tree to somewhere.

In other words, ->s_vfs_rename_mutex is sufficient to stabilize "X is an
ancestor of Y" only if X and Y are already in the same tree.  Otherwise
it can go from false to true, and one can construct a deadlock on that.

Make lock_two_directories() report an error in such case and update the
callers of lock_rename()/lock_rename_child() to handle such errors.

And yes, such conditions are not impossible to create ;-/

Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>

1 files changed, 4 insertions, 0 deletions

diff --git a/fs/overlayfs/dir.c b/fs/overlayfs/dir.c
index aab3f5d93556..0f8b4a719237 100644
--- a/fs/overlayfs/dir.c
+++ b/fs/overlayfs/dir.c
@@ -1180,6 +1180,10 @@ static int ovl_rename(struct mnt_idmap *idmap, struct inode *olddir,
 	}
 
 	trap = lock_rename(new_upperdir, old_upperdir);
+	if (IS_ERR(trap)) {
+		err = PTR_ERR(trap);
+		goto out_revert_creds;
+	}
 
 	olddentry = ovl_lookup_upper(ofs, old->d_name.name, old_upperdir,
 				     old->d_name.len);