switch bugfixes

CHANGELOG.md

@@ -43,6 +43,7 @@ and this project adheres to [Semantic Versioning](http://semver.org/spec/v2.0.0.
 ### Fixed
 - Memory leaks in the DHT fixed
 - Crash where ICMPv6 NDP goroutine would incorrectly start in TUN mode fixed
+- Remove peers from the switch table of they stop sending switch messages but keep the TCP connection alive
 
 ## [0.2.7] - 2018-10-13
 ### Added

src/yggdrasil/switch.go

@@ -208,6 +208,7 @@ func (t *switchTable) doMaintenance() {
 	defer t.mutex.Unlock() // Release lock when we're done
 	t.cleanRoot()
 	t.cleanDropped()
+	t.cleanPeers()
 }
 
 // Updates the root periodically if it is ourself, or promotes ourself to root if we're better than the current root or if the current root has timed out.
@@ -258,11 +259,33 @@ func (t *switchTable) forgetPeer(port switchPort) {
 	if port != t.parent {
 		return
 	}
+	t.parent = 0
 	for _, info := range t.data.peers {
 		t.unlockedHandleMsg(&info.msg, info.port, true)
 	}
 }
 
+// Clean all unresponsive peers from the table, needed in case a peer stops updating.
+// Needed in case a non-parent peer keeps the connection open but stops sending updates.
+// Also reclaims space from deleted peers by copying the map.
+func (t switchTable) cleanPeers() {
+	now := time.Now()
+	for port, peer := range t.data.peers {
+		if now.Sub(peer.time) > switch_timeout+switch_throttle {
+			// Longer than switch_timeout to make sure we don't remove a working peer because the root stopped responding.
+			delete(t.data.peers, port)
+		}
+	}
+	if _, isIn := t.data.peers[t.parent]; !isIn {
+		// The root timestamp would probably time out before this happens, but better safe than sorry.
+		// We removed the current parent, so find a new one.
+		t.parent = 0
+		for _, peer := range t.data.peers {
+			t.unlockedHandleMsg(&peer.msg, peer.port, true)
+		}
+	}
+}
+
 // Dropped is a list of roots that are better than the current root, but stopped sending new timestamps.
 // If we switch to a new root, and that root is better than an old root that previously timed out, then we can clean up the old dropped root infos.
 // This function is called periodically to do that cleanup.
@@ -379,9 +402,6 @@ func (t *switchTable) unlockedHandleMsg(msg *switchMsg, fromPort switchPort, rep
 	if !equiv(&sender.locator, &oldSender.locator) {
 		// Reset faster info, we'll start refilling it right after this
 		sender.faster = nil
-		for _, peer := range t.data.peers {
-			delete(peer.faster, sender.port)
-		}
 		doUpdate = true
 	}
 	// Update the matrix of peer "faster" thresholds