ceph: return error for traceless reply race

When we receives traceless reply for request that created new inode,
we re-send a lookup request to MDS get information of the newly created
inode. (VFS expects FS' callback return an inode in create case)
This breaks one request into two requests. Other client may modify or
move to the new inode in the middle.

When the race happens, ceph_handle_notrace_create() unconditionally
links the dentry for 'create' operation to the inode returned by lookup.
This may confuse VFS when the inode is a directory (VFS does not allow
multiple linkages for directory inode).

This patch makes ceph_handle_notrace_create() when it detect a race.
This event should be rare and it happens only when we talk to old MDS.
Recent MDS does not send traceless reply for request that creates new
inode.

Signed-off-by: Yan, Zheng <zyan@redhat.com>

fs/ceph/dir.c

@@ -670,14 +670,17 @@ int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
 		/*
 		 * We created the item, then did a lookup, and found
 		 * it was already linked to another inode we already
-		 * had in our cache (and thus got spliced).  Link our
-		 * dentry to that inode, but don't hash it, just in
-		 * case the VFS wants to dereference it.
+		 * had in our cache (and thus got spliced). To not
+		 * confuse VFS (especially when inode is a directory),
+		 * we don't link our dentry to that inode, return an
+		 * error instead.
+		 *
+		 * This event should be rare and it happens only when
+		 * we talk to old MDS. Recent MDS does not send traceless
+		 * reply for request that creates new inode.
 		 */
-		BUG_ON(!result->d_inode);
-		d_instantiate(dentry, result->d_inode);
 		d_drop(result);
-		return 0;
+		return -ESTALE;
 	}
 	return PTR_ERR(result);
 }