gossipd: reply to large listchannels in parts.

This has two effects: most importantly, it avoids the problem where
lightningd creates a 800MB JSON blob in response to listchannels,
which causes OOM on the Raspberry Pi (our previous max allocation was
832MB).  This is because lightning-cli can start draining the JSON
while we're filling the buffer, so we end up with a max allocation of
68MB.

But despite being less efficient (multiple queries to gossipd), it
actually speeds things up due to the parallelism:

MCP with -O3 -flto before vs after:
-listchannels_sec:8.980000-9.330000(9.206+/-0.14)
+listchannels_sec:7.500000-7.830000(7.656+/-0.11)

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

gossipd/gossip_wire.csv

@@ -45,8 +45,10 @@ gossip_getroute_reply,,hops,num_hops*struct route_hop
 gossip_getchannels_request,3007
 gossip_getchannels_request,,short_channel_id,?struct short_channel_id
 gossip_getchannels_request,,source,?struct node_id
+gossip_getchannels_request,,prev,?struct short_channel_id
 
 gossip_getchannels_reply,3107
+gossip_getchannels_reply,,complete,bool
 gossip_getchannels_reply,,num_channels,u32
 gossip_getchannels_reply,,nodes,num_channels*struct gossip_getchannels_entry
 

gossipd/gossipd.c

@@ -2133,11 +2133,13 @@ static struct io_plan *getchannels_req(struct io_conn *conn,
 	u8 *out;
 	const struct gossip_getchannels_entry **entries;
 	struct chan *chan;
-	struct short_channel_id *scid;
+	struct short_channel_id *scid, *prev;
 	struct node_id *source;
+	bool complete = true;
 
 	/* Note: scid is marked optional in gossip_wire.csv */
-	if (!fromwire_gossip_getchannels_request(msg, msg, &scid, &source))
+	if (!fromwire_gossip_getchannels_request(msg, msg, &scid, &source,
+						 &prev))
 		master_badmsg(WIRE_GOSSIP_GETCHANNELS_REQUEST, msg);
 
 	entries = tal_arr(tmpctx, const struct gossip_getchannels_entry *, 0);
@@ -2161,15 +2163,20 @@ static struct io_plan *getchannels_req(struct io_conn *conn,
 		u64 idx;
 
 		/* For the more general case, we just iterate through every
-		 * short channel id. */
-		for (chan = uintmap_first(&daemon->rstate->chanmap, &idx);
-		     chan;
-		     chan = uintmap_after(&daemon->rstate->chanmap, &idx)) {
+		 * short channel id, starting with previous if any (there is
+		 * no scid 0). */
+		idx = prev ? prev->u64 : 0;
+		while ((chan = uintmap_after(&daemon->rstate->chanmap, &idx))) {
 			append_channel(daemon->rstate, &entries, chan, NULL);
+			/* Limit how many we do at once. */
+			if (tal_count(entries) == 4096) {
+				complete = false;
+				break;
+			}
 		}
 	}
 
-	out = towire_gossip_getchannels_reply(NULL, entries);
+	out = towire_gossip_getchannels_reply(NULL, complete, entries);
 	daemon_conn_send(daemon->master, take(out));
 	return daemon_conn_read_next(conn, daemon->master);
 }

lightningd/gossip_control.c

@@ -390,30 +390,51 @@ static void json_add_halfchan(struct json_stream *response,
 	json_object_end(response);
 }
 
+struct listchannels_info {
+	struct command *cmd;
+	struct json_stream *response;
+	struct short_channel_id *id;
+	struct node_id *source;
+};
+
 /* Called upon receiving a getchannels_reply from `gossipd` */
 static void json_listchannels_reply(struct subd *gossip UNUSED, const u8 *reply,
-				   const int *fds UNUSED, struct command *cmd)
+				    const int *fds UNUSED,
+				    struct listchannels_info *linfo)
 {
 	size_t i;
 	struct gossip_getchannels_entry **entries;
-	struct json_stream *response;
+	bool complete;
 
-	if (!fromwire_gossip_getchannels_reply(reply, reply, &entries)) {
-		was_pending(command_fail(cmd, LIGHTNINGD,
-					 "Invalid reply from gossipd"));
+	if (!fromwire_gossip_getchannels_reply(reply, reply,
+					       &complete, &entries)) {
+		/* Shouldn't happen: just end json stream. */
+		log_broken(linfo->cmd->ld->log, "Invalid reply from gossipd");
+		was_pending(command_raw_complete(linfo->cmd, linfo->response));
 		return;
 	}
 
-	response = json_stream_success(cmd);
-	json_object_start(response, NULL);
-	json_array_start(response, "channels");
 	for (i = 0; i < tal_count(entries); i++) {
-		json_add_halfchan(response, entries[i], 0);
-		json_add_halfchan(response, entries[i], 1);
+		json_add_halfchan(linfo->response, entries[i], 0);
+		json_add_halfchan(linfo->response, entries[i], 1);
+	}
+
+	/* More coming?  Ask from this point on.. */
+	if (!complete) {
+		u8 *req;
+		assert(tal_count(entries) != 0);
+		req = towire_gossip_getchannels_request(linfo->cmd,
+							linfo->id,
+							linfo->source,
+							&entries[i-1]
+							->short_channel_id);
+		subd_req(linfo->cmd->ld->gossip, linfo->cmd->ld->gossip,
+			 req, -1, 0, json_listchannels_reply, linfo);
+	} else {
+		json_array_end(linfo->response);
+		json_object_end(linfo->response);
+		was_pending(command_success(linfo->cmd, linfo->response));
 	}
-	json_array_end(response);
-	json_object_end(response);
-	was_pending(command_success(cmd, response));
 }
 
 static struct command_result *json_listchannels(struct command *cmd,
@@ -422,21 +443,29 @@ static struct command_result *json_listchannels(struct command *cmd,
 						const jsmntok_t *params)
 {
 	u8 *req;
-	struct short_channel_id *id;
-	struct node_id *source;
+	struct listchannels_info *linfo = tal(cmd, struct listchannels_info);
 
+	linfo->cmd = cmd;
 	if (!param(cmd, buffer, params,
-		   p_opt("short_channel_id", param_short_channel_id, &id),
-		   p_opt("source", param_node_id, &source),
+		   p_opt("short_channel_id", param_short_channel_id, &linfo->id),
+		   p_opt("source", param_node_id, &linfo->source),
 		   NULL))
 		return command_param_failed();
 
-	if (id && source)
+	if (linfo->id && linfo->source)
 		return command_fail(cmd, JSONRPC2_INVALID_PARAMS,
 				    "Cannot specify both source and short_channel_id");
-	req = towire_gossip_getchannels_request(cmd, id, source);
+
+	/* Start JSON response, then we stream. */
+	linfo->response = json_stream_success(cmd);
+	json_object_start(linfo->response, NULL);
+	json_array_start(linfo->response, "channels");
+
+	req = towire_gossip_getchannels_request(cmd, linfo->id, linfo->source,
+						NULL);
 	subd_req(cmd->ld->gossip, cmd->ld->gossip,
-		 req, -1, 0, json_listchannels_reply, cmd);
+		 req, -1, 0, json_listchannels_reply, linfo);
+
 	return command_still_pending(cmd);
 }