gossipd: fail peer for the master daemon.

This fixes the only case where the master currently has to write directly
to the peer: re-sending an error.  We make gossipd do it, by adding
a new gossipctl_fail_peer message.

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

gossipd/gossip.c

@@ -99,14 +99,15 @@ static void destroy_peer(struct peer *peer)
 	}
 }
 
-static struct peer *setup_new_peer(struct daemon *daemon, const u8 *msg)
+static struct peer *setup_new_peer(struct daemon *daemon,
+				   u64 unique_id,
+				   const struct crypto_state *cs)
 {
 	struct peer *peer = tal(daemon, struct peer);
 
 	init_peer_crypto_state(peer, &peer->pcs);
-	if (!fromwire_gossipctl_new_peer(msg, NULL, &peer->unique_id,
-					 &peer->pcs.cs))
-		return tal_free(peer);
+	peer->pcs.cs = *cs;
+	peer->unique_id = unique_id;
 	peer->daemon = daemon;
 	peer->error = NULL;
 	peer->local = true;
@@ -115,6 +116,7 @@ static struct peer *setup_new_peer(struct daemon *daemon, const u8 *msg)
 	msg_queue_init(&peer->peer_out, peer);
 	list_add_tail(&daemon->peers, &peer->list);
 	tal_add_destructor(peer, destroy_peer);
+
 	wake_pkt_out(peer);
 	return peer;
 }
@@ -416,10 +418,14 @@ static struct io_plan *new_peer_got_fd(struct io_conn *conn, struct peer *peer)
 static struct io_plan *new_peer(struct io_conn *conn, struct daemon *daemon,
 				const u8 *msg)
 {
-	struct peer *peer = setup_new_peer(daemon, msg);
-	if (!peer)
-		status_failed(STATUS_FAIL_MASTER_IO,
-			      "bad gossipctl_new_peer: %s", tal_hex(trc, msg));
+	struct peer *peer;
+	struct crypto_state cs;
+	u64 unique_id;
+
+	if (!fromwire_gossipctl_new_peer(msg, NULL, &unique_id, &cs))
+		master_badmsg(WIRE_GOSSIPCTL_NEW_PEER, msg);
+
+	peer = setup_new_peer(daemon, unique_id, &cs);
 	return io_recv_fd(conn, &peer->fd, new_peer_got_fd, peer);
 }
 
@@ -433,6 +439,45 @@ static struct peer *find_peer(struct daemon *daemon, u64 unique_id)
 	return NULL;
 }
 
+/* We send error, then close. */
+static struct io_plan *peer_send_error(struct io_conn *conn, struct peer *peer)
+{
+	const u8 *out = msg_dequeue(&peer->peer_out);
+	return peer_write_message(conn, &peer->pcs, take(out),
+				  (void *)io_close_cb);
+}
+
+static struct io_plan *fail_peer_got_fd(struct io_conn *conn, struct peer *peer)
+{
+	peer->conn = io_new_conn(conn, peer->fd, peer_send_error, peer);
+	if (!peer->conn) {
+		peer->error = "Could not create connection";
+		tal_free(peer);
+	} else {
+		/* If conn dies, we forget peer. */
+		tal_steal(peer->conn, peer);
+	}
+	return daemon_conn_read_next(conn, &peer->daemon->master);
+}
+
+static struct io_plan *fail_peer(struct io_conn *conn, struct daemon *daemon,
+				 const u8 *msg)
+{
+	struct peer *peer;
+	struct crypto_state cs;
+	u64 unique_id;
+	u8 *failmsg;
+
+	if (!fromwire_gossipctl_fail_peer(msg, msg, NULL, &unique_id, &cs,
+					  &failmsg))
+		master_badmsg(WIRE_GOSSIPCTL_FAIL_PEER, msg);
+
+	peer = setup_new_peer(daemon, unique_id, &cs);
+	msg_enqueue(&peer->peer_out, take(failmsg));
+
+	return io_recv_fd(conn, &peer->fd, fail_peer_got_fd, peer);
+}
+
 static struct io_plan *release_peer(struct io_conn *conn, struct daemon *daemon,
 				    const u8 *msg)
 {
@@ -742,6 +787,10 @@ static struct io_plan *recv_req(struct io_conn *conn, struct daemon_conn *master
 	case WIRE_GOSSIP_FORWARDED_MSG:
 		handle_forwarded_msg(conn, daemon, daemon->master.msg_in);
 		return daemon_conn_read_next(conn, &daemon->master);
+
+	case WIRE_GOSSIPCTL_FAIL_PEER:
+		return fail_peer(conn, daemon, master->msg_in);
+
 	case WIRE_GOSSIPCTL_RELEASE_PEER_REPLY:
 	case WIRE_GOSSIPCTL_RELEASE_PEER_REPLYFAIL:
 	case WIRE_GOSSIPCTL_GET_PEER_GOSSIPFD_REPLY:

gossipd/gossip_wire.csv

@@ -18,7 +18,6 @@ gossipctl_init,,broadcast_interval,u32
 gossipctl_init,,chain_hash,struct sha256_double
 
 # These take an fd, but have no response
-# (if it is to move onto a channel, we get a status msg).
 gossipctl_new_peer,3001
 gossipctl_new_peer,,unique_id,u64
 gossipctl_new_peer,,crypto_state,struct crypto_state
@@ -112,3 +111,10 @@ gossipctl_get_peer_gossipfd_reply,3112
 
 # Failure (can't make new socket)
 gossipctl_get_peer_gossipfd_replyfail,3212
+
+# Send canned message to peer and fail it.
+gossipctl_fail_peer,3013
+gossipctl_fail_peer,,unique_id,u64
+gossipctl_fail_peer,,crypto_state,struct crypto_state
+gossipctl_fail_peer,,len,u16
+gossipctl_fail_peer,,failmsg,len*u8

lightningd/Makefile

@@ -18,7 +18,6 @@ LIGHTNINGD_COMMON_OBJS :=			\
 	common/channel_config.o			\
 	common/configdir.o			\
 	common/crypto_state.o			\
-	common/cryptomsg.o			\
 	common/derive_basepoints.o		\
 	common/funding_tx.o			\
 	common/htlc_state.o			\

lightningd/gossip_control.c

@@ -95,6 +95,7 @@ static int gossip_msg(struct subd *gossip, const u8 *msg, const int *fds)
 	/* These are messages we send, not them. */
 	case WIRE_GOSSIPCTL_INIT:
 	case WIRE_GOSSIPCTL_NEW_PEER:
+	case WIRE_GOSSIPCTL_FAIL_PEER:
 	case WIRE_GOSSIPCTL_RELEASE_PEER:
 	case WIRE_GOSSIPCTL_DROP_PEER:
 	case WIRE_GOSSIPCTL_GET_PEER_GOSSIPFD:

lightningd/peer_control.c

@@ -262,34 +262,6 @@ static void peer_start_closingd(struct peer *peer,
 				int peer_fd, int gossip_fd,
 				bool reconnected);
 
-/* FIXME: Fake NOP dev_disconnect/dev_sabotage_fd for below. */
-enum dev_disconnect dev_disconnect(int pkt_type)
-{
-	return DEV_DISCONNECT_NORMAL;
-}
-
-void dev_sabotage_fd(int fd)
-{
-	abort();
-}
-
-void dev_blackhole_fd(int fd)
-{
-	abort();
-}
-
-/* Send (encrypted) error message, then close. */
-static struct io_plan *send_error(struct io_conn *conn,
-				  struct peer_crypto_state *pcs)
-{
-	log_debug(pcs->peer->log, "Sending canned error");
-	/* FIXME: This is the only place where master talks directly to peer;
-	 * and pulls in quite a lot of code to do so.  If we got a subdaemon
-	 * to do this work, we'd avoid pulling in cryptomsg.o and the fake
-	 * dev_disconnect. */
-	return peer_write_message(conn, pcs, pcs->peer->error, (void *)io_close_cb);
-}
-
 /* FIXME:
  *
  * This is a lot of code duplication!  We should turn gossipd into welcomed(?):
@@ -483,10 +455,14 @@ static bool peer_reconnected(struct lightningd *ld,
 	 * retransmit the error packet and ignore any other packets for that
 	 * channel, and the following requirements do not apply. */
 	if (peer->error) {
-		struct peer_crypto_state *pcs = tal(peer, struct peer_crypto_state);
-		init_peer_crypto_state(peer, pcs);
-		pcs->cs = *cs;
-		tal_steal(io_new_conn(peer, fd, send_error, pcs), pcs);
+		/* FIXME: we should do this in response to reestablish, unless
+		 * global error */
+		log_debug(peer->log, "Sending canned error");
+		subd_send_msg(peer->ld->gossip,
+			      take(towire_gossipctl_fail_peer(peer,
+							      peer->unique_id,
+							      cs, peer->error)));
+		subd_send_fd(peer->ld->gossip, fd);
 		return true;
 	}