ceph: clear dir complete, invalidate dentry on replayed rename
authorSage Weil <sage@newdream.net>
Wed, 28 Apr 2010 23:12:06 +0000 (16:12 -0700)
committerSage Weil <sage@newdream.net>
Mon, 3 May 2010 17:49:25 +0000 (10:49 -0700)
If a rename operation is resent to the MDS following an MDS restart, the
client does not get a full reply (containing the resulting metadata) back.
In that case, a ceph_rename() needs to compensate by doing anything useful
that fill_inode() would have, like d_move().

It also needs to invalidate the dentry (to workaround the vfs_rename_dir()
bug) and clear the dir complete flag, just like fill_trace().

Signed-off-by: Sage Weil <sage@newdream.net>
fs/ceph/dir.c

index ea8ee2e526aae94d5fd10c16cf4e78ca16090818..650d2db5ed26a48923266fd2bc836d92e4edf1f1 100644 (file)
@@ -880,7 +880,16 @@ static int ceph_rename(struct inode *old_dir, struct dentry *old_dentry,
                 * do_request, above).  If there is no trace, we need
                 * to do it here.
                 */
+
+               /* d_move screws up d_subdirs order */
+               ceph_i_clear(new_dir, CEPH_I_COMPLETE);
+
                d_move(old_dentry, new_dentry);
+
+               /* ensure target dentry is invalidated, despite
+                  rehashing bug in vfs_rename_dir */
+               new_dentry->d_time = jiffies;
+               ceph_dentry(new_dentry)->lease_shared_gen = 0;
        }
        ceph_mdsc_put_request(req);
        return err;