ceph: fix readdir vs fragmentation race

following sequence of events tigger the race

- client readdir frag 0* -> got item 'A'
- MDS merges frag 0* and frag 1*
- client send readdir request (frag 1*, offset 2, readdir_start 'A')
- MDS reply items (that are after item 'A') in frag *

Link: http://tracker.ceph.com/issues/17286Signed-off-by: Yan, Zheng <zyan@redhat.com>

fs/ceph/inode.c

@@ -1511,7 +1511,8 @@ int ceph_readdir_prepopulate(struct ceph_mds_request *req,
 			ceph_fill_dirfrag(d_inode(parent), rinfo->dir_dir);
 	}
 
-	if (ceph_frag_is_leftmost(frag) && req->r_readdir_offset == 2) {
+	if (ceph_frag_is_leftmost(frag) && req->r_readdir_offset == 2 &&
+	    !(rinfo->hash_order && req->r_path2)) {
 		/* note dir version at start of readdir so we can tell
 		 * if any dentries get dropped */
 		req->r_dir_release_cnt = atomic64_read(&ci->i_release_count);