]> git.bitcoin.ninja Git - rust-lightning/commitdiff
Optimize ChannelMonitor persistence on block connections.
authorG8XSU <3442979+G8XSU@users.noreply.github.com>
Fri, 14 Jun 2024 23:56:36 +0000 (16:56 -0700)
committerG8XSU <3442979+G8XSU@users.noreply.github.com>
Wed, 19 Jun 2024 07:04:23 +0000 (00:04 -0700)
Currently, every block connection triggers the persistence of all
ChannelMonitors with an updated best_block. This approach poses
challenges for large node operators managing thousands of channels.
Furthermore, it leads to a thundering herd problem
(https://en.wikipedia.org/wiki/Thundering_herd_problem), overwhelming
the storage with simultaneous requests.

To address this issue, we now persist ChannelMonitors at a
regular cadence, spreading their persistence across blocks to
mitigate spikes in write operations.

Outcome: After doing this, Ldk's IO footprint should be reduced
by ~50 times. The processing time required to sync each block
will be significantly reduced, particularly for nodes with 1000s
of channels, as write latency plays a significant role in this process.
As a result, the Node/ChainMonitor will be blocked for a shorter
duration, leading to further efficiency gains.

lightning/src/chain/chainmonitor.rs
lightning/src/chain/channelmonitor.rs
lightning/src/chain/onchaintx.rs

index 58bf5f8002d27cc7ab092fc8cc288a8e52377acf..e6bb9d90778ce46b4cd7ba5f25eafb52daba455d 100644 (file)
@@ -44,6 +44,7 @@ use crate::prelude::*;
 use crate::sync::{RwLock, RwLockReadGuard, Mutex, MutexGuard};
 use core::ops::Deref;
 use core::sync::atomic::{AtomicUsize, Ordering};
+use bitcoin::hashes::Hash;
 use bitcoin::secp256k1::PublicKey;
 
 /// `Persist` defines behavior for persisting channel monitors: this could mean
@@ -260,10 +261,11 @@ where C::Target: chain::Filter,
        {
                let err_str = "ChannelMonitor[Update] persistence failed unrecoverably. This indicates we cannot continue normal operation and must shut down.";
                let funding_outpoints = hash_set_from_iter(self.monitors.read().unwrap().keys().cloned());
+               let channel_count = funding_outpoints.len();
                for funding_outpoint in funding_outpoints.iter() {
                        let monitor_lock = self.monitors.read().unwrap();
                        if let Some(monitor_state) = monitor_lock.get(funding_outpoint) {
-                               if self.update_monitor_with_chain_data(header, txdata, &process, funding_outpoint, &monitor_state).is_err() {
+                               if self.update_monitor_with_chain_data(header, best_height, txdata, &process, funding_outpoint, &monitor_state, channel_count).is_err() {
                                        // Take the monitors lock for writing so that we poison it and any future
                                        // operations going forward fail immediately.
                                        core::mem::drop(monitor_lock);
@@ -278,7 +280,7 @@ where C::Target: chain::Filter,
                let monitor_states = self.monitors.write().unwrap();
                for (funding_outpoint, monitor_state) in monitor_states.iter() {
                        if !funding_outpoints.contains(funding_outpoint) {
-                               if self.update_monitor_with_chain_data(header, txdata, &process, funding_outpoint, &monitor_state).is_err() {
+                               if self.update_monitor_with_chain_data(header, best_height, txdata, &process, funding_outpoint, &monitor_state, channel_count).is_err() {
                                        log_error!(self.logger, "{}", err_str);
                                        panic!("{}", err_str);
                                }
@@ -297,14 +299,29 @@ where C::Target: chain::Filter,
        }
 
        fn update_monitor_with_chain_data<FN>(
-               &self, header: &Header, txdata: &TransactionData, process: FN, funding_outpoint: &OutPoint,
-               monitor_state: &MonitorHolder<ChannelSigner>
+               &self, header: &Header, best_height: Option<u32>, txdata: &TransactionData, process: FN, funding_outpoint: &OutPoint,
+               monitor_state: &MonitorHolder<ChannelSigner>, channel_count: usize,
        ) -> Result<(), ()> where FN: Fn(&ChannelMonitor<ChannelSigner>, &TransactionData) -> Vec<TransactionOutputs> {
                let monitor = &monitor_state.monitor;
                let logger = WithChannelMonitor::from(&self.logger, &monitor, None);
-               let mut txn_outputs;
-               {
-                       txn_outputs = process(monitor, txdata);
+
+               let mut txn_outputs = process(monitor, txdata);
+
+               let get_partition_key = |funding_outpoint: &OutPoint| {
+                       let funding_txid_hash = funding_outpoint.txid.to_raw_hash();
+                       let funding_txid_hash_bytes = funding_txid_hash.as_byte_array();
+                       let funding_txid_u32 = u32::from_be_bytes([funding_txid_hash_bytes[0], funding_txid_hash_bytes[1], funding_txid_hash_bytes[2], funding_txid_hash_bytes[3]]);
+                       funding_txid_u32.wrapping_add(best_height.unwrap_or_default())
+               };
+
+               let partition_factor = if channel_count < 15 {
+                       5
+               } else {
+                       50 // ~ 8hours
+               };
+
+               let has_pending_claims = monitor_state.monitor.has_pending_claims();
+               if has_pending_claims || get_partition_key(funding_outpoint) % partition_factor == 0 {
                        log_trace!(logger, "Syncing Channel Monitor for channel {}", log_funding_info!(monitor));
                        match self.persister.update_persisted_channel(*funding_outpoint, None, monitor) {
                                ChannelMonitorUpdateStatus::Completed =>
@@ -313,10 +330,10 @@ where C::Target: chain::Filter,
                                        ),
                                ChannelMonitorUpdateStatus::InProgress => {
                                        log_trace!(logger, "Channel Monitor sync for channel {} in progress.", log_funding_info!(monitor));
-                               },
+                               }
                                ChannelMonitorUpdateStatus::UnrecoverableError => {
                                        return Err(());
-                               },
+                               }
                        }
                }
 
@@ -870,14 +887,17 @@ impl<ChannelSigner: EcdsaChannelSigner, C: Deref, T: Deref, F: Deref, L: Deref,
 
 #[cfg(test)]
 mod tests {
-       use crate::check_added_monitors;
+       use crate::{check_added_monitors, check_closed_event};
        use crate::{expect_payment_path_successful, get_event_msg};
        use crate::{get_htlc_update_msgs, get_revoke_commit_msgs};
        use crate::chain::{ChannelMonitorUpdateStatus, Watch};
-       use crate::events::{Event, MessageSendEvent, MessageSendEventsProvider};
+       use crate::chain::channelmonitor::ANTI_REORG_DELAY;
+       use crate::events::{ClosureReason, Event, MessageSendEvent, MessageSendEventsProvider};
        use crate::ln::functional_test_utils::*;
        use crate::ln::msgs::ChannelMessageHandler;
 
+       const CHAINSYNC_MONITOR_PARTITION_FACTOR: u32 = 5;
+
        #[test]
        fn test_async_ooo_offchain_updates() {
                // Test that if we have multiple offchain updates being persisted and they complete
@@ -983,6 +1003,79 @@ mod tests {
                check_added_monitors!(nodes[0], 1);
        }
 
+       #[test]
+       fn test_chainsync_triggers_distributed_monitor_persistence() {
+               let chanmon_cfgs = create_chanmon_cfgs(3);
+               let node_cfgs = create_node_cfgs(3, &chanmon_cfgs);
+               let node_chanmgrs = create_node_chanmgrs(3, &node_cfgs, &[None, None, None]);
+               let nodes = create_network(3, &node_cfgs, &node_chanmgrs);
+
+               // Use FullBlockViaListen to avoid duplicate calls to process_chain_data and skips_blocks() in
+               // case of other connect_styles.
+               *nodes[0].connect_style.borrow_mut() = ConnectStyle::FullBlockViaListen;
+               *nodes[1].connect_style.borrow_mut() = ConnectStyle::FullBlockViaListen;
+               *nodes[2].connect_style.borrow_mut() = ConnectStyle::FullBlockViaListen;
+
+               let _channel_1 = create_announced_chan_between_nodes(&nodes, 0, 1).2;
+               let channel_2 = create_announced_chan_between_nodes_with_value(&nodes, 0, 2, 1_000_000, 0).2;
+
+               chanmon_cfgs[0].persister.chain_sync_monitor_persistences.lock().unwrap().clear();
+               chanmon_cfgs[1].persister.chain_sync_monitor_persistences.lock().unwrap().clear();
+               chanmon_cfgs[2].persister.chain_sync_monitor_persistences.lock().unwrap().clear();
+
+               connect_blocks(&nodes[0], CHAINSYNC_MONITOR_PARTITION_FACTOR * 2);
+               connect_blocks(&nodes[1], CHAINSYNC_MONITOR_PARTITION_FACTOR * 2);
+               connect_blocks(&nodes[2], CHAINSYNC_MONITOR_PARTITION_FACTOR * 2);
+
+               // Connecting [`DEFAULT_CHAINSYNC_PARTITION_FACTOR`] * 2 blocks should trigger only 2 writes
+               // per monitor/channel.
+               assert_eq!(2 * 2, chanmon_cfgs[0].persister.chain_sync_monitor_persistences.lock().unwrap().len());
+               assert_eq!(2, chanmon_cfgs[1].persister.chain_sync_monitor_persistences.lock().unwrap().len());
+               assert_eq!(2, chanmon_cfgs[2].persister.chain_sync_monitor_persistences.lock().unwrap().len());
+
+               // Test that monitors with pending_claims are persisted on every block.
+               // Now, close channel_2 i.e. b/w node-0 and node-2 to create pending_claim in node[0].
+               nodes[0].node.force_close_broadcasting_latest_txn(&channel_2, &nodes[2].node.get_our_node_id(), "Channel force-closed".to_string()).unwrap();
+               check_closed_event!(&nodes[0], 1, ClosureReason::HolderForceClosed { broadcasted_latest_txn: Some(true) }, false,
+                       [nodes[2].node.get_our_node_id()], 1000000);
+               check_closed_broadcast(&nodes[0], 1, true);
+               let close_tx = nodes[0].tx_broadcaster.txn_broadcasted.lock().unwrap().split_off(0);
+               assert_eq!(close_tx.len(), 1);
+
+               mine_transaction(&nodes[2], &close_tx[0]);
+               check_added_monitors(&nodes[2], 1);
+               check_closed_broadcast(&nodes[2], 1, true);
+               check_closed_event!(&nodes[2], 1, ClosureReason::CommitmentTxConfirmed, false,
+                       [nodes[0].node.get_our_node_id()], 1000000);
+
+               chanmon_cfgs[0].persister.chain_sync_monitor_persistences.lock().unwrap().clear();
+               chanmon_cfgs[2].persister.chain_sync_monitor_persistences.lock().unwrap().clear();
+
+               // For channel_2, there should be a monitor write for every block connection.
+               // We connect [`DEFAULT_CHAINSYNC_MONITOR_PARTITION_FACTOR`] blocks since we don't know when
+               // channel_1 monitor persistence will occur, with [`DEFAULT_CHAINSYNC_MONITOR_PARTITION_FACTOR`]
+               // it will be persisted exactly once.
+               connect_blocks(&nodes[0], CHAINSYNC_MONITOR_PARTITION_FACTOR);
+               connect_blocks(&nodes[2], CHAINSYNC_MONITOR_PARTITION_FACTOR);
+
+               // DEFAULT_CHAINSYNC_MONITOR_PARTITION_FACTOR writes for channel_2 due to pending_claim, 1 for
+               // channel_1
+               assert_eq!((CHAINSYNC_MONITOR_PARTITION_FACTOR + 1) as usize, chanmon_cfgs[0].persister.chain_sync_monitor_persistences.lock().unwrap().len());
+               // For node[2], there is no pending_claim
+               assert_eq!(1, chanmon_cfgs[2].persister.chain_sync_monitor_persistences.lock().unwrap().len());
+
+               // Confirm claim for node[0] with ANTI_REORG_DELAY and reset monitor write counter.
+               mine_transaction(&nodes[0], &close_tx[0]);
+               connect_blocks(&nodes[0], ANTI_REORG_DELAY - 1);
+               check_added_monitors(&nodes[0], 1);
+               chanmon_cfgs[0].persister.chain_sync_monitor_persistences.lock().unwrap().clear();
+
+               // Again connect 1 full cycle of DEFAULT_CHAINSYNC_MONITOR_PARTITION_FACTOR blocks, it should only
+               // result in 1 write per monitor/channel.
+               connect_blocks(&nodes[0], CHAINSYNC_MONITOR_PARTITION_FACTOR);
+               assert_eq!(2, chanmon_cfgs[0].persister.chain_sync_monitor_persistences.lock().unwrap().len());
+       }
+
        #[test]
        #[cfg(feature = "std")]
        fn update_during_chainsync_poisons_channel() {
@@ -991,12 +1084,15 @@ mod tests {
                let node_chanmgrs = create_node_chanmgrs(2, &node_cfgs, &[None, None]);
                let nodes = create_network(2, &node_cfgs, &node_chanmgrs);
                create_announced_chan_between_nodes(&nodes, 0, 1);
+               *nodes[0].connect_style.borrow_mut() = ConnectStyle::FullBlockViaListen;
 
                chanmon_cfgs[0].persister.set_update_ret(ChannelMonitorUpdateStatus::UnrecoverableError);
 
                assert!(std::panic::catch_unwind(|| {
                        // Returning an UnrecoverableError should always panic immediately
-                       connect_blocks(&nodes[0], 1);
+                       // Connecting [`DEFAULT_CHAINSYNC_PARTITION_FACTOR`] blocks so that we trigger some persistence
+                       // after accounting for block-height based partitioning/distribution.
+                       connect_blocks(&nodes[0], CHAINSYNC_MONITOR_PARTITION_FACTOR);
                }).is_err());
                assert!(std::panic::catch_unwind(|| {
                        // ...and also poison our locks causing later use to panic as well
index cd6061a20307650e87d126afab9cc0f65aa1bbf5..c6d1bcdd2158025fa11984aa32f5207380d459cd 100644 (file)
@@ -1812,6 +1812,12 @@ impl<Signer: EcdsaChannelSigner> ChannelMonitor<Signer> {
                );
        }
 
+       /// Returns true if the monitor has pending claim requests that are not fully confirmed yet.
+       pub fn has_pending_claims(&self) -> bool
+       {
+               self.inner.lock().unwrap().onchain_tx_handler.has_pending_claims()
+       }
+
        /// Triggers rebroadcasts of pending claims from a force-closed channel after a transaction
        /// signature generation failure.
        pub fn signer_unblocked<B: Deref, F: Deref, L: Deref>(
index e5757cfd1e52f5b1da358f4fd88a706052447147..07a75361d67678a8c7011e18c09c805817e5e53d 100644 (file)
@@ -535,6 +535,13 @@ impl<ChannelSigner: EcdsaChannelSigner> OnchainTxHandler<ChannelSigner> {
                }
        }
 
+       /// Returns true if we are currently tracking any pending claim requests that are not fully 
+       /// confirmed yet.
+       pub(super) fn has_pending_claims(&self) -> bool
+       {
+               self.pending_claim_requests.len() != 0
+       }
+
        /// Lightning security model (i.e being able to redeem/timeout HTLC or penalize counterparty
        /// onchain) lays on the assumption of claim transactions getting confirmed before timelock
        /// expiration (CSV or CLTV following cases). In case of high-fee spikes, claim tx may get stuck