Merge pull request #10012 from ziggie1984/fix-goroutine-leak

multi: prevent goroutine leak in brontide

discovery/gossiper.go

@@ -411,6 +411,10 @@ type processedNetworkMsg struct {
 
 // cachedNetworkMsg is a wrapper around a network message that can be used with
 // *lru.Cache.
+//
+// NOTE: This struct is not thread safe which means you need to assure no
+// concurrent read write access to it and all its contents which are pointers
+// as well.
 type cachedNetworkMsg struct {
 	msgs []*processedNetworkMsg
 }
@@ -3131,7 +3135,9 @@ func (d *AuthenticatedGossiper) handleChanUpdate(ctx context.Context,
 
 		// NOTE: We don't return anything on the error channel for this
 		// message, as we expect that will be done when this
-		// ChannelUpdate is later reprocessed.
+		// ChannelUpdate is later reprocessed. This might never happen
+		// if the corresponding ChannelAnnouncement is never received
+		// or the LRU cache is filled up and the entry is evicted.
 		return nil, false
 
 	default:

docs/release-notes/release-notes-0.19.2.md

@@ -32,6 +32,10 @@
 - [Fixed](https://github.com/lightningnetwork/lnd/pull/9978) a deadlock which
   can happen when the peer start-up has not yet completed but a another p2p
   connection attempt tries to disconnect the peer.
+  
+- [Fixed](https://github.com/lightningnetwork/lnd/pull/10012) a case which
+  could lead to a memory issues due to a goroutine leak in the peer/gossiper
+  code.
 
 # New Features
 

peer/brontide.go

@@ -1990,32 +1990,22 @@ func newDiscMsgStream(p *Brontide) *msgStream {
 		// so that a parent context can be passed in here.
 		ctx := context.TODO()
 
+		// Processing here means we send it to the gossiper which then
+		// decides whether this message is processed immediately or
+		// waits for dependent messages to be processed. It can also
+		// happen that the message is not processed at all if it is
+		// premature and the LRU cache fills up and the message is
+		// deleted.
 		p.log.Debugf("Processing remote msg %T", msg)
 
-		errChan := p.cfg.AuthGossiper.ProcessRemoteAnnouncement(
+		// TODO(ziggie): ProcessRemoteAnnouncement returns an error
-			ctx, msg, p,
+		// channel, but we cannot rely on it being written to.
-		)
+		// Because some messages might never be processed (e.g.
-
+		// premature channel updates). We should change the design here
-		// Start a goroutine to process the error channel for logging
+		// and use the actor model pattern as soon as it is available.
-		// purposes.
+		// So for now we should NOT use the error channel.
-		//
+		// See https://github.com/lightningnetwork/lnd/pull/9820.
-		// TODO(ziggie): Maybe use the error to potentially punish the
+		p.cfg.AuthGossiper.ProcessRemoteAnnouncement(ctx, msg, p)
-		// peer depending on the error ?
-		go func() {
-			select {
-			case <-p.cg.Done():
-				return
-
-			case err := <-errChan:
-				if err != nil {
-					p.log.Warnf("Error processing remote "+
-						"msg %T: %v", msg,
-						err)
-				}
-			}
-
-			p.log.Debugf("Processed remote msg %T", msg)
-		}()
 	}
 
 	return newMsgStream(