lockd: fix lockd shutdown race

As reported by David Jeffery: "a signal was sent to lockd while lockd
was shutting down from a request to stop nfs.  The signal causes lockd
to call restart_grace() which puts the lockd_net structure on the grace
list.  If this signal is received at the wrong time, it will occur after
lockd_down_net() has called locks_end_grace() but before
lockd_down_net() stops the lockd thread.  This leads to lockd putting
the lockd_net structure back on the grace list, then exiting without
anything removing it from the list."

So, perform the final locks_end_grace() from the the lockd thread; this
ensures it's serialized with respect to restart_grace().
Reported-by: David Jeffery <djeffery@redhat.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

fs/lockd/svc.c

@@ -132,6 +132,8 @@ lockd(void *vrqstp)
 {
 	int		err = 0;
 	struct svc_rqst *rqstp = vrqstp;
+	struct net *net = &init_net;
+	struct lockd_net *ln = net_generic(net, lockd_net_id);
 
 	/* try_to_freeze() is called from svc_recv() */
 	set_freezable();
@@ -176,6 +178,8 @@ lockd(void *vrqstp)
 	if (nlmsvc_ops)
 		nlmsvc_invalidate_all();
 	nlm_shutdown_hosts();
+	cancel_delayed_work_sync(&ln->grace_period_end);
+	locks_end_grace(&ln->lockd_manager);
 	return 0;
 }
 
@@ -270,8 +274,6 @@ static void lockd_down_net(struct svc_serv *serv, struct net *net)
 	if (ln->nlmsvc_users) {
 		if (--ln->nlmsvc_users == 0) {
 			nlm_shutdown_hosts_net(net);
-			cancel_delayed_work_sync(&ln->grace_period_end);
-			locks_end_grace(&ln->lockd_manager);
 			svc_shutdown_net(serv, net);
 			dprintk("lockd_down_net: per-net data destroyed; net=%p\n", net);
 		}