p2p/dnsdisc: fix hot-spin when all trees are empty

p2p/dnsdisc/client.go

@@ -238,10 +238,10 @@ func (it *randomIterator) Node() *enode.Node {
 
 // Close closes the iterator.
 func (it *randomIterator) Close() {
+	it.cancelFn()
+
 	it.mu.Lock()
 	defer it.mu.Unlock()
-
-	it.cancelFn()
 	it.trees = nil
 }
 
@@ -264,7 +264,7 @@ func (it *randomIterator) addTree(url string) error {
 // nextNode syncs random tree entries until it finds a node.
 func (it *randomIterator) nextNode() *enode.Node {
 	for {
-		ct := it.nextTree()
+		ct := it.pickTree()
 		if ct == nil {
 			return nil
 		}
@@ -282,26 +282,68 @@ func (it *randomIterator) nextNode() *enode.Node {
 	}
 }
 
-// nextTree returns a random tree.
-func (it *randomIterator) nextTree() *clientTree {
+// pickTree returns a random tree to sync from.
+func (it *randomIterator) pickTree() *clientTree {
 	it.mu.Lock()
 	defer it.mu.Unlock()
 
 	if it.lc.changed {
 		it.rebuildTrees()
 		it.lc.changed = false
 	}
-	if len(it.trees) == 0 {
-		return nil
+
+	for {
+		// Find trees that might still have pending items to sync.
+		// If there are any, pick a random syncable tree.
+		syncable, disabled := it.syncableTrees()
+		if len(syncable) > 0 {
+			return syncable[rand.Intn(len(syncable))]
+		}
+		if len(disabled) == 0 {
+			return nil // Iterator was closed.
+		}
+		// No sync action can be performed on any tree right now. The only meaningful
+		// thing to do is waiting for any root record to get updated.
+		if !it.waitForRootUpdates(disabled) {
+			return nil // Iterator was closed.
+		}
 	}
-	limit := rand.Intn(len(it.trees))
+}
+
+// syncableTrees finds trees on which any meaningful sync action can be performed.
+func (it *randomIterator) syncableTrees() (syncable, disabled []*clientTree) {
 	for _, ct := range it.trees {
-		if limit == 0 {
-			return ct
+		if ct.canSyncRandom() {
+			syncable = append(syncable, ct)
+		} else {
+			disabled = append(disabled, ct)
 		}
-		limit--
 	}
-	return nil
+	return syncable, disabled
+}
+
+// waitForRootUpdates waits for the closest scheduled root check time on the given trees.
+func (it *randomIterator) waitForRootUpdates(trees []*clientTree) bool {
+	var checkTree string
+	var nextCheck = mclock.AbsTime(-1)
+	for _, ct := range trees {
+		check := ct.nextScheduledRootCheck()
+		if nextCheck == -1 || check < nextCheck {
+			nextCheck = check
+			checkTree = ct.loc.domain
+		}
+	}
+
+	sleep := nextCheck.Sub(it.c.clock.Now())
+	it.c.cfg.Logger.Debug("DNS iterator waiting for root updates", "sleep", sleep, "tree", checkTree)
+	timeout := it.c.clock.NewTimer(sleep)
+	defer timeout.Stop()
+	select {
+	case <-timeout.C():
+		return true
+	case <-it.ctx.Done():
+		return false // Iterator was closed.
+	}
 }
 
 // rebuildTrees rebuilds the 'trees' map.

p2p/dnsdisc/client_test.go

@@ -231,6 +231,53 @@ func TestIteratorRootRecheckOnFail(t *testing.T) {
 	checkIterator(t, it, nodes)
 }
 
+// This test checks that the iterator works correctly when the tree is initially empty.
+func TestIteratorEmptyTree(t *testing.T) {
+	var (
+		clock    = new(mclock.Simulated)
+		nodes    = testNodes(nodesSeed1, 1)
+		resolver = newMapResolver()
+		c        = NewClient(Config{
+			Resolver:        resolver,
+			Logger:          testlog.Logger(t, log.LvlTrace),
+			RecheckInterval: 20 * time.Minute,
+			RateLimit:       500,
+		})
+	)
+	c.clock = clock
+	tree1, url := makeTestTree("n", nil, nil)
+	tree2, _ := makeTestTree("n", nodes, nil)
+	resolver.add(tree1.ToTXT("n"))
+
+	// Start the iterator.
+	node := make(chan *enode.Node)
+	it, err := c.NewIterator(url)
+	if err != nil {
+		t.Fatal(err)
+	}
+	go func() {
+		it.Next()
+		node <- it.Node()
+	}()
+
+	// Wait for the client to get stuck in waitForRootUpdates.
+	clock.WaitForTimers(1)
+
+	// Now update the root.
+	resolver.add(tree2.ToTXT("n"))
+
+	// Wait for it to pick up the root change.
+	clock.Run(c.cfg.RecheckInterval)
+	select {
+	case n := <-node:
+		if n.ID() != nodes[0].ID() {
+			t.Fatalf("wrong node returned")
+		}
+	case <-time.After(5 * time.Second):
+		t.Fatal("it.Next() did not unblock within 5s of real time")
+	}
+}
+
 // updateSomeNodes applies ENR updates to some of the given nodes.
 func updateSomeNodes(keySeed int64, nodes []*enode.Node) {
 	keys := testKeys(nodesSeed1, len(nodes))

p2p/dnsdisc/sync.go

@@ -25,9 +25,9 @@ import (
 	"github.com/ethereum/go-ethereum/p2p/enode"
 )
 
-const (
-	rootRecheckFailCount = 5 // update root if this many leaf requests fail
-)
+// This is the number of consecutive leaf requests that may fail before
+// we consider re-resolving the tree root.
+const rootRecheckFailCount = 5
 
 // clientTree is a full tree being synced.
 type clientTree struct {
@@ -89,13 +89,22 @@ func (ct *clientTree) syncRandom(ctx context.Context) (n *enode.Node, err error)
 	ct.gcLinks()
 
 	// Sync next random entry in ENR tree. Once every node has been visited, we simply
-	// start over. This is fine because entries are cached.
+	// start over. This is fine because entries are cached internally by the client LRU
+	// also by DNS resolvers.
 	if ct.enrs.done() {
 		ct.enrs = newSubtreeSync(ct.c, ct.loc, ct.root.eroot, false)
 	}
 	return ct.syncNextRandomENR(ctx)
 }
 
+// canSyncRandom checks if any meaningful action can be performed by syncRandom.
+func (ct *clientTree) canSyncRandom() bool {
+	// Note: the check for non-zero leaf count is very important here.
+	// If we're done syncing all nodes, and no leaves were found, the tree
+	// is empty and we can't use it for sync.
+	return ct.rootUpdateDue() || !ct.links.done() || !ct.enrs.done() || ct.enrs.leaves != 0
+}
+
 // gcLinks removes outdated links from the global link cache. GC runs once
 // when the link sync finishes.
 func (ct *clientTree) gcLinks() {
@@ -184,10 +193,14 @@ func (ct *clientTree) updateRoot(ctx context.Context) error {
 // rootUpdateDue returns true when a root update is needed.
 func (ct *clientTree) rootUpdateDue() bool {
 	tooManyFailures := ct.leafFailCount > rootRecheckFailCount
-	scheduledCheck := ct.c.clock.Now().Sub(ct.lastRootCheck) > ct.c.cfg.RecheckInterval
+	scheduledCheck := ct.c.clock.Now() >= ct.nextScheduledRootCheck()
 	return ct.root == nil || tooManyFailures || scheduledCheck
 }
 
+func (ct *clientTree) nextScheduledRootCheck() mclock.AbsTime {
+	return ct.lastRootCheck.Add(ct.c.cfg.RecheckInterval)
+}
+
 // slowdownRootUpdate applies a delay to root resolution if is tried
 // too frequently. This avoids busy polling when the client is offline.
 // Returns true if the timeout passed, false if sync was canceled.
@@ -218,10 +231,11 @@ type subtreeSync struct {
 	root    string
 	missing []string // missing tree node hashes
 	link    bool     // true if this sync is for the link tree
+	leaves  int      // counter of synced leaves
 }
 
 func newSubtreeSync(c *Client, loc *linkEntry, root string, link bool) *subtreeSync {
-	return &subtreeSync{c, loc, root, []string{root}, link}
+	return &subtreeSync{c, loc, root, []string{root}, link, 0}
 }
 
 func (ts *subtreeSync) done() bool {
@@ -253,10 +267,12 @@ func (ts *subtreeSync) resolveNext(ctx context.Context, hash string) (entry, err
 		if ts.link {
 			return nil, errENRInLinkTree
 		}
+		ts.leaves++
 	case *linkEntry:
 		if !ts.link {
 			return nil, errLinkInENRTree
 		}
+		ts.leaves++
 	case *branchEntry:
 		ts.missing = append(ts.missing, e.children...)
 	}