Return an error if we can't connect and limit the number of retries

gossipd/gossip.c

@@ -100,6 +100,15 @@ struct reaching {
 
 	/* Did we succeed? */
 	bool succeeded;
+
+	/* How many times have we attempted to connect? */
+	u32 attempts;
+
+	/* How many times to attempt */
+	u32 max_attempts;
+
+	/* Timestamp of the first attempt */
+	u32 first_attempt;
 };
 
 /* Things we need when we're talking direct to the peer. */
@@ -1557,13 +1566,25 @@ static void try_connect(struct reaching *reach);
 
 static void connect_failed(struct io_conn *conn, struct reaching *reach)
 {
-	status_trace("Failed connected out for %s, will try again",
-		     type_to_string(trc, struct pubkey, &reach->id));
-
-	/* FIXME: Configurable timer! */
-	new_reltimer(&reach->daemon->timers, reach,
-		     time_from_sec(5),
-		     try_connect, reach);
+	u32 diff = time_now().ts.tv_sec - reach->first_attempt;
+	reach->attempts++;
+
+	if (reach->attempts >= reach->max_attempts) {
+		status_trace("Failed to connect after %d attempts, giving up "
+			     "after %d seconds",
+			     reach->attempts, diff);
+		daemon_conn_send(
+		    &reach->daemon->master,
+		    take(towire_gossip_peer_connection_failed(
+			conn, &reach->id, diff, reach->attempts, false)));
+		tal_free(reach);
+	} else {
+		status_trace("Failed connected out for %s, will try again",
+			     type_to_string(trc, struct pubkey, &reach->id));
+		/* FIXME: Configurable timer! */
+		new_reltimer(&reach->daemon->timers, reach, time_from_sec(5),
+			     try_connect, reach);
+	}
 }
 
 static struct io_plan *conn_init(struct io_conn *conn, struct reaching *reach)
@@ -1625,6 +1646,12 @@ static void try_connect(struct reaching *reach)
 		/* FIXME: add reach_failed message */
 		status_trace("No address known for %s, giving up",
 			     type_to_string(trc, struct pubkey, &reach->id));
+		daemon_conn_send(
+		    &reach->daemon->master,
+		    take(towire_gossip_peer_connection_failed(
+			reach, &reach->id,
+			time_now().ts.tv_sec - reach->first_attempt,
+			reach->attempts, true)));
 		tal_free(reach);
 		return;
 	}
@@ -1683,6 +1710,9 @@ static bool try_reach_peer(struct daemon *daemon, const struct pubkey *id)
 	reach->succeeded = false;
 	reach->daemon = daemon;
 	reach->id = *id;
+	reach->first_attempt = time_now().ts.tv_sec;
+	reach->attempts = 0;
+	reach->max_attempts = 10;
 	list_add_tail(&daemon->reaching, &reach->list);
 	tal_add_destructor(reach, destroy_reaching);
 
@@ -1913,6 +1943,7 @@ static struct io_plan *recv_req(struct io_conn *conn, struct daemon_conn *master
 	case WIRE_GOSSIP_RESOLVE_CHANNEL_REPLY:
 	case WIRE_GOSSIP_PEER_CONNECTED:
 	case WIRE_GOSSIP_PEER_ALREADY_CONNECTED:
+	case WIRE_GOSSIP_PEER_CONNECTION_FAILED:
 	case WIRE_GOSSIP_PEER_NONGOSSIP:
 	case WIRE_GOSSIP_GET_UPDATE:
 	case WIRE_GOSSIP_GET_UPDATE_REPLY:

gossipd/gossip_wire.csv

@@ -43,6 +43,13 @@ gossip_peer_connected,,lfeatures,lflen*u8
 gossip_peer_already_connected,3015
 gossip_peer_already_connected,,id,struct pubkey
 
+# gossipd -> master: attempted to connect, unsuccessful, gave up
+gossip_peer_connection_failed,3020
+gossip_peer_connection_failed,,id,struct pubkey
+gossip_peer_connection_failed,,timeout,u32
+gossip_peer_connection_failed,,attempts,u32
+gossip_peer_connection_failed,,addr_unknown,bool
+
 # Gossipd -> master: peer sent non-gossip packet. Two fds: peer and gossip
 gossip_peer_nongossip,3003
 gossip_peer_nongossip,,id,struct pubkey

lightningd/gossip_control.c

@@ -133,6 +133,9 @@ static unsigned gossip_msg(struct subd *gossip, const u8 *msg, const int *fds)
 	case WIRE_GOSSIP_PEER_ALREADY_CONNECTED:
 		peer_already_connected(gossip->ld, msg);
 		break;
+	case WIRE_GOSSIP_PEER_CONNECTION_FAILED:
+		peer_connection_failed(gossip->ld, msg);
+		break;
 	case WIRE_GOSSIP_PEER_NONGOSSIP:
 		if (tal_count(fds) != 2)
 			return 2;

lightningd/peer_control.c

@@ -606,6 +606,39 @@ void peer_already_connected(struct lightningd *ld, const u8 *msg)
 	connect_succeeded(ld, &id);
 }
 
+void peer_connection_failed(struct lightningd *ld, const u8 *msg)
+{
+	struct pubkey id;
+	u32 attempts, timediff;
+	struct connect *i, *next;
+	bool addr_unknown;
+	char *error;
+
+	if (!fromwire_gossip_peer_connection_failed(msg, NULL, &id, &attempts,
+						    &timediff, &addr_unknown))
+		fatal(
+		    "Gossip gave bad GOSSIP_PEER_CONNECTION_FAILED message %s",
+		    tal_hex(msg, msg));
+
+	if (addr_unknown) {
+		error = tal_fmt(
+		    msg, "No address known for node %s, please provide one",
+		    type_to_string(msg, struct pubkey, &id));
+	} else {
+		error = tal_fmt(msg, "Could not connect to %s after %d seconds and %d attempts",
+				type_to_string(msg, struct pubkey, &id), timediff,
+				attempts);
+	}
+
+	/* Careful!  Completing command frees connect. */
+	list_for_each_safe(&ld->connects, i, next, list) {
+		if (!pubkey_eq(&i->id, &id))
+			continue;
+
+		command_fail(i->cmd, "%s", error);
+	}
+}
+
 void peer_sent_nongossip(struct lightningd *ld,
 			 const struct pubkey *id,
 			 const struct wireaddr *addr,

lightningd/peer_control.h

@@ -179,6 +179,9 @@ void peer_connected(struct lightningd *ld, const u8 *msg,
 /* This simply means we asked to reach a peer, but we already have it */
 void peer_already_connected(struct lightningd *ld, const u8 *msg);
 
+/* We were unable to connect to the peer */
+void peer_connection_failed(struct lightningd *ld, const u8 *msg);
+
 void peer_sent_nongossip(struct lightningd *ld,
 			 const struct pubkey *id,
 			 const struct wireaddr *addr,