Fix gossip using `gossip_timestamp_filter` instead of queries 2022-03-gossip-queries-sucks
authorMatt Corallo <git@bluematt.me>
Thu, 24 Mar 2022 03:00:06 +0000 (03:00 +0000)
committerMatt Corallo <git@bluematt.me>
Fri, 25 Mar 2022 17:12:09 +0000 (17:12 +0000)
See large comment added for more details

lightning/src/routing/network_graph.rs

index 012357df361767e330a698f492846a6b44e8917b..5268d443771cf48d668c1dee6e3a05c14e1c6b39 100644 (file)
@@ -401,82 +401,91 @@ where C::Target: chain::Access, L::Target: Logger
                        return ();
                }
 
-               // Send a gossip_timestamp_filter to enable gossip message receipt. Note that we have to
-               // use a "all timestamps" filter as sending the current timestamp would result in missing
-               // gossip messages that are simply sent late. We could calculate the intended filter time
-               // by looking at the current time and subtracting two weeks (before which we'll reject
-               // messages), but there's not a lot of reason to bother - our peers should be discarding
-               // the same messages.
+               // The lightning network's gossip sync system is completely broken in numerous ways.
+               //
+               // Given no broadly-available set-reconciliation protocol, the only reasonable approach is
+               // to do a full sync from the first few peers we connect to, and then receive gossip
+               // updates from all our peers normally.
+               //
+               // Originally, we could simply tell a peer to dump us the entire gossip table on startup,
+               // wasting lots of bandwidth but ensuring we have the full network graph. After the initial
+               // dump peers would always send gossip and we'd stay up-to-date with whatever our peer has
+               // seen.
+               //
+               // In order to reduce the bandwidth waste, "gossip queries" were introduced, allowing you
+               // to ask for the SCIDs of all channels in your peer's routing graph, and then only request
+               // channel data which you are missing. Except there was no way at all to identify which
+               // `channel_update`s you were missing, so you still had to request everything, just in a
+               // very complicated way with some queries instead of just getting the dump.
+               //
+               // Later, an option was added to fetch the latest timestamps of the `channel_update`s to
+               // make efficient sync possible, however it has yet to be implemented in lnd, which makes
+               // relying on it useless.
+               //
+               // After gossip queries were introduced, support for receiving a full gossip table dump on
+               // connection was removed from several nodes, making it impossible to get a full sync
+               // without using the "gossip queries" messages.
+               //
+               // Once you opt into "gossip queries" the only way to receive any gossip updates that a
+               // peer receives after you connect, you must send a `gossip_timestamp_filter` message. This
+               // message, as the name implies, tells the peer to not forward any gossip messages with a
+               // timestamp older than a given value (not the time the peer received the filter, but the
+               // timestamp in the update message, which is often hours behind when the peer received the
+               // message).
+               //
+               // Obnoxiously, `gossip_timestamp_filter` isn't *just* a filter, but its also a request for
+               // your peer to send you the full routing graph (subject to the filter). Thus, in order to
+               // tell a peer to send you any updates as it sees them, you have to also ask for the full
+               // routing graph to be synced. If you set a timestamp filter near the current time, peers
+               // will simply not forward any new updates they see to you which were generated some time
+               // ago (which is not uncommon). If you instead set a timestamp filter near 0 (or two weeks
+               // ago), you will always get the full routing graph from all your peers.
+               //
+               // Most lightning nodes today opt to simply turn off receiving gossip data which only
+               // propagated some time after it was generated, and, worse, often disable gossiping with
+               // several peers after their first connection. The second behavior can cause gossip to not
+               // propagate fully if there are cuts in the gossiping subgraph.
+               //
+               // In an attempt to cut a middle ground between always fetching the full graph from all of
+               // our peers and never receiving gossip from peers at all, we send all of our peers a
+               // `gossip_timestamp_filter`, with the filter time set either two weeks ago or an hour ago.
+               //
+               // For no-std builds, we bury our head in the sand and do a full sync on each connection.
+               let should_request_full_sync = self.should_request_full_sync(&their_node_id);
+               #[allow(unused_mut, unused_assignments)]
+               let mut gossip_start_time = 0;
+               #[cfg(feature = "std")]
+               {
+                       gossip_start_time = SystemTime::now().duration_since(UNIX_EPOCH).expect("Time must be > 1970").as_secs();
+                       if should_request_full_sync {
+                               gossip_start_time -= 60 * 60 * 24 * 7 * 2; // 2 weeks ago
+                       } else {
+                               gossip_start_time -= 60 * 60; // an hour ago
+                       }
+               }
+
                let mut pending_events = self.pending_events.lock().unwrap();
                pending_events.push(MessageSendEvent::SendGossipTimestampFilter {
                        node_id: their_node_id.clone(),
                        msg: GossipTimestampFilter {
                                chain_hash: self.network_graph.genesis_hash,
-                               first_timestamp: 0,
+                               first_timestamp: gossip_start_time as u32, // 2106 issue!
                                timestamp_range: u32::max_value(),
                        },
                });
-
-               // Check if we need to perform a full synchronization with this peer
-               if !self.should_request_full_sync(&their_node_id) {
-                       return ();
-               }
-
-               let first_blocknum = 0;
-               let number_of_blocks = 0xffffffff;
-               log_debug!(self.logger, "Sending query_channel_range peer={}, first_blocknum={}, number_of_blocks={}", log_pubkey!(their_node_id), first_blocknum, number_of_blocks);
-               pending_events.push(MessageSendEvent::SendChannelRangeQuery {
-                       node_id: their_node_id.clone(),
-                       msg: QueryChannelRange {
-                               chain_hash: self.network_graph.genesis_hash,
-                               first_blocknum,
-                               number_of_blocks,
-                       },
-               });
        }
 
-       /// Statelessly processes a reply to a channel range query by immediately
-       /// sending an SCID query with SCIDs in the reply. To keep this handler
-       /// stateless, it does not validate the sequencing of replies for multi-
-       /// reply ranges. It does not validate whether the reply(ies) cover the
-       /// queried range. It also does not filter SCIDs to only those in the
-       /// original query range. We also do not validate that the chain_hash
-       /// matches the chain_hash of the NetworkGraph. Any chan_ann message that
-       /// does not match our chain_hash will be rejected when the announcement is
-       /// processed.
-       fn handle_reply_channel_range(&self, their_node_id: &PublicKey, msg: ReplyChannelRange) -> Result<(), LightningError> {
-               log_debug!(self.logger, "Handling reply_channel_range peer={}, first_blocknum={}, number_of_blocks={}, sync_complete={}, scids={}", log_pubkey!(their_node_id), msg.first_blocknum, msg.number_of_blocks, msg.sync_complete, msg.short_channel_ids.len(),);
-
-               log_debug!(self.logger, "Sending query_short_channel_ids peer={}, batch_size={}", log_pubkey!(their_node_id), msg.short_channel_ids.len());
-               let mut pending_events = self.pending_events.lock().unwrap();
-               pending_events.push(MessageSendEvent::SendShortIdsQuery {
-                       node_id: their_node_id.clone(),
-                       msg: QueryShortChannelIds {
-                               chain_hash: msg.chain_hash,
-                               short_channel_ids: msg.short_channel_ids,
-                       }
-               });
-
+       fn handle_reply_channel_range(&self, _their_node_id: &PublicKey, _msg: ReplyChannelRange) -> Result<(), LightningError> {
+               // We don't make queries, so should never receive replies. If, in the future, the set
+               // reconciliation extensions to gossip queries become broadly supported, we should revert
+               // this code to its state pre-0.0.106.
                Ok(())
        }
 
-       /// When an SCID query is initiated the remote peer will begin streaming
-       /// gossip messages. In the event of a failure, we may have received
-       /// some channel information. Before trying with another peer, the
-       /// caller should update its set of SCIDs that need to be queried.
-       fn handle_reply_short_channel_ids_end(&self, their_node_id: &PublicKey, msg: ReplyShortChannelIdsEnd) -> Result<(), LightningError> {
-               log_debug!(self.logger, "Handling reply_short_channel_ids_end peer={}, full_information={}", log_pubkey!(their_node_id), msg.full_information);
-
-               // If the remote node does not have up-to-date information for the
-               // chain_hash they will set full_information=false. We can fail
-               // the result and try again with a different peer.
-               if !msg.full_information {
-                       return Err(LightningError {
-                               err: String::from("Received reply_short_channel_ids_end with no information"),
-                               action: ErrorAction::IgnoreError
-                       });
-               }
-
+       fn handle_reply_short_channel_ids_end(&self, _their_node_id: &PublicKey, _msg: ReplyShortChannelIdsEnd) -> Result<(), LightningError> {
+               // We don't make queries, so should never receive replies. If, in the future, the set
+               // reconciliation extensions to gossip queries become broadly supported, we should revert
+               // this code to its state pre-0.0.106.
                Ok(())
        }
 
@@ -1535,7 +1544,7 @@ mod tests {
        use routing::network_graph::{NetGraphMsgHandler, NetworkGraph, NetworkUpdate, MAX_EXCESS_BYTES_FOR_RELAY};
        use ln::msgs::{Init, OptionalField, RoutingMessageHandler, UnsignedNodeAnnouncement, NodeAnnouncement,
                UnsignedChannelAnnouncement, ChannelAnnouncement, UnsignedChannelUpdate, ChannelUpdate, 
-               ReplyChannelRange, ReplyShortChannelIdsEnd, QueryChannelRange, QueryShortChannelIds, MAX_VALUE_MSAT};
+               ReplyChannelRange, QueryChannelRange, QueryShortChannelIds, MAX_VALUE_MSAT};
        use util::test_utils;
        use util::logger::Logger;
        use util::ser::{Readable, Writeable};
@@ -2272,15 +2281,16 @@ mod tests {
        }
 
        #[test]
+       #[cfg(feature = "std")]
        fn calling_sync_routing_table() {
+               use std::time::{SystemTime, UNIX_EPOCH};
+
                let network_graph = create_network_graph();
                let (secp_ctx, net_graph_msg_handler) = create_net_graph_msg_handler(&network_graph);
                let node_privkey_1 = &SecretKey::from_slice(&[42; 32]).unwrap();
                let node_id_1 = PublicKey::from_secret_key(&secp_ctx, node_privkey_1);
 
                let chain_hash = genesis_block(Network::Testnet).header.block_hash();
-               let first_blocknum = 0;
-               let number_of_blocks = 0xffff_ffff;
 
                // It should ignore if gossip_queries feature is not enabled
                {
@@ -2290,132 +2300,23 @@ mod tests {
                        assert_eq!(events.len(), 0);
                }
 
-               // It should send a query_channel_message with the correct information
+               // It should send a gossip_timestamp_filter with the correct information
                {
                        let init_msg = Init { features: InitFeatures::known(), remote_network_address: None };
                        net_graph_msg_handler.peer_connected(&node_id_1, &init_msg);
                        let events = net_graph_msg_handler.get_and_clear_pending_msg_events();
-                       assert_eq!(events.len(), 2);
+                       assert_eq!(events.len(), 1);
                        match &events[0] {
                                MessageSendEvent::SendGossipTimestampFilter{ node_id, msg } => {
                                        assert_eq!(node_id, &node_id_1);
                                        assert_eq!(msg.chain_hash, chain_hash);
-                                       assert_eq!(msg.first_timestamp, 0);
+                                       let expected_timestamp = SystemTime::now().duration_since(UNIX_EPOCH).expect("Time must be > 1970").as_secs();
+                                       assert!((msg.first_timestamp as u64) >= expected_timestamp - 60*60*24*7*2);
+                                       assert!((msg.first_timestamp as u64) < expected_timestamp - 60*60*24*7*2 + 10);
                                        assert_eq!(msg.timestamp_range, u32::max_value());
                                },
                                _ => panic!("Expected MessageSendEvent::SendChannelRangeQuery")
                        };
-                       match &events[1] {
-                               MessageSendEvent::SendChannelRangeQuery{ node_id, msg } => {
-                                       assert_eq!(node_id, &node_id_1);
-                                       assert_eq!(msg.chain_hash, chain_hash);
-                                       assert_eq!(msg.first_blocknum, first_blocknum);
-                                       assert_eq!(msg.number_of_blocks, number_of_blocks);
-                               },
-                               _ => panic!("Expected MessageSendEvent::SendChannelRangeQuery")
-                       };
-               }
-
-               // It should not enqueue a query when should_request_full_sync return false.
-               // The initial implementation allows syncing with the first 5 peers after
-               // which should_request_full_sync will return false
-               {
-                       let network_graph = create_network_graph();
-                       let (secp_ctx, net_graph_msg_handler) = create_net_graph_msg_handler(&network_graph);
-                       let init_msg = Init { features: InitFeatures::known(), remote_network_address: None };
-                       for n in 1..7 {
-                               let node_privkey = &SecretKey::from_slice(&[n; 32]).unwrap();
-                               let node_id = PublicKey::from_secret_key(&secp_ctx, node_privkey);
-                               net_graph_msg_handler.peer_connected(&node_id, &init_msg);
-                               let events = net_graph_msg_handler.get_and_clear_pending_msg_events();
-                               if n <= 5 {
-                                       assert_eq!(events.len(), 2);
-                               } else {
-                                       // Even after the we stop sending the explicit query, we should still send a
-                                       // gossip_timestamp_filter on each new connection.
-                                       assert_eq!(events.len(), 1);
-                               }
-
-                       }
-               }
-       }
-
-       #[test]
-       fn handling_reply_channel_range() {
-               let network_graph = create_network_graph();
-               let (secp_ctx, net_graph_msg_handler) = create_net_graph_msg_handler(&network_graph);
-               let node_privkey_1 = &SecretKey::from_slice(&[42; 32]).unwrap();
-               let node_id_1 = PublicKey::from_secret_key(&secp_ctx, node_privkey_1);
-
-               let chain_hash = genesis_block(Network::Testnet).header.block_hash();
-
-               // Test receipt of a single reply that should enqueue an SCID query
-               // matching the SCIDs in the reply
-               {
-                       let result = net_graph_msg_handler.handle_reply_channel_range(&node_id_1, ReplyChannelRange {
-                               chain_hash,
-                               sync_complete: true,
-                               first_blocknum: 0,
-                               number_of_blocks: 2000,
-                               short_channel_ids: vec![
-                                       0x0003e0_000000_0000, // 992x0x0
-                                       0x0003e8_000000_0000, // 1000x0x0
-                                       0x0003e9_000000_0000, // 1001x0x0
-                                       0x0003f0_000000_0000, // 1008x0x0
-                                       0x00044c_000000_0000, // 1100x0x0
-                                       0x0006e0_000000_0000, // 1760x0x0
-                               ],
-                       });
-                       assert!(result.is_ok());
-
-                       // We expect to emit a query_short_channel_ids message with the received scids
-                       let events = net_graph_msg_handler.get_and_clear_pending_msg_events();
-                       assert_eq!(events.len(), 1);
-                       match &events[0] {
-                               MessageSendEvent::SendShortIdsQuery { node_id, msg } => {
-                                       assert_eq!(node_id, &node_id_1);
-                                       assert_eq!(msg.chain_hash, chain_hash);
-                                       assert_eq!(msg.short_channel_ids, vec![
-                                               0x0003e0_000000_0000, // 992x0x0
-                                               0x0003e8_000000_0000, // 1000x0x0
-                                               0x0003e9_000000_0000, // 1001x0x0
-                                               0x0003f0_000000_0000, // 1008x0x0
-                                               0x00044c_000000_0000, // 1100x0x0
-                                               0x0006e0_000000_0000, // 1760x0x0
-                                       ]);
-                               },
-                               _ => panic!("expected MessageSendEvent::SendShortIdsQuery"),
-                       }
-               }
-       }
-
-       #[test]
-       fn handling_reply_short_channel_ids() {
-               let network_graph = create_network_graph();
-               let (secp_ctx, net_graph_msg_handler) = create_net_graph_msg_handler(&network_graph);
-               let node_privkey = &SecretKey::from_slice(&[41; 32]).unwrap();
-               let node_id = PublicKey::from_secret_key(&secp_ctx, node_privkey);
-
-               let chain_hash = genesis_block(Network::Testnet).header.block_hash();
-
-               // Test receipt of a successful reply
-               {
-                       let result = net_graph_msg_handler.handle_reply_short_channel_ids_end(&node_id, ReplyShortChannelIdsEnd {
-                               chain_hash,
-                               full_information: true,
-                       });
-                       assert!(result.is_ok());
-               }
-
-               // Test receipt of a reply that indicates the peer does not maintain up-to-date information
-               // for the chain_hash requested in the query.
-               {
-                       let result = net_graph_msg_handler.handle_reply_short_channel_ids_end(&node_id, ReplyShortChannelIdsEnd {
-                               chain_hash,
-                               full_information: false,
-                       });
-                       assert!(result.is_err());
-                       assert_eq!(result.err().unwrap().err, "Received reply_short_channel_ids_end with no information");
                }
        }