diff --git a/p2p/discover/common.go b/p2p/discover/common.go
index cbd62f867d..193c6c12cc 100644
--- a/p2p/discover/common.go
+++ b/p2p/discover/common.go
@@ -17,8 +17,8 @@
 package discover
 
 import (
-	"context"
 	"crypto/ecdsa"
+	"math/rand"
 	"net"
 	"sync"
 
@@ -64,22 +64,27 @@ type lookupFunc func(func(*enode.Node))
 // lookupWalker performs recursive lookups, walking the DHT.
 // It manages a set iterators which receive lookup results as they are found.
 type lookupWalker struct {
-	lookup lookupFunc
-
 	newIterCh chan *lookupIterator
 	delIterCh chan *lookupIterator
 	triggerCh chan struct{}
 	closeCh   chan struct{}
 	wg        sync.WaitGroup
+
+	lookup     lookupFunc
+	lookupDone chan struct{}
+	foundNode  chan *enode.Node
+	iters      map[*lookupIterator]struct{}
 }
 
 func newLookupWalker(fn lookupFunc) *lookupWalker {
 	w := &lookupWalker{
 		lookup:    fn,
+		iters:     make(map[*lookupIterator]struct{}),
 		newIterCh: make(chan *lookupIterator),
 		delIterCh: make(chan *lookupIterator),
 		triggerCh: make(chan struct{}),
 		closeCh:   make(chan struct{}),
+		foundNode: make(chan *enode.Node),
 	}
 	w.wg.Add(1)
 	go w.loop()
@@ -92,38 +97,33 @@ func (w *lookupWalker) close() {
 }
 
 func (w *lookupWalker) loop() {
-	var (
-		iters      = make(map[*lookupIterator]struct{})
-		foundNode  = make(chan *enode.Node)
-		lookupDone = make(chan struct{}, 1)
-		trigger    = w.triggerCh
-	)
 	for {
 		select {
 		case it := <-w.newIterCh:
-			iters[it] = struct{}{}
+			w.iters[it] = struct{}{}
+			w.startNewLookup()
 
 		case it := <-w.delIterCh:
-			delete(iters, it)
+			delete(w.iters, it)
 
-		case <-trigger:
-			trigger = nil // stop listening to trigger until lookupDone
-			go w.runLookup(foundNode, lookupDone)
+		case <-w.triggerCh:
+			w.startNewLookup()
 
-		case <-lookupDone:
-			trigger = w.triggerCh
+		case <-w.lookupDone:
+			w.lookupDone = nil
+			w.startNewLookup()
 
-		case n := <-foundNode:
-			for it := range iters {
+		case n := <-w.foundNode:
+			for it := range w.iters {
 				it.deliver(n)
 			}
 
 		case <-w.closeCh:
-			for it := range iters {
-				it.drainAndClose()
+			for it := range w.iters {
+				it.close()
 			}
-			if trigger == nil {
-				<-lookupDone
+			if w.lookupDone != nil {
+				<-w.lookupDone
 			}
 			w.wg.Done()
 			return
@@ -131,85 +131,112 @@ func (w *lookupWalker) loop() {
 	}
 }
 
-func (w *lookupWalker) runLookup(nodes chan<- *enode.Node, done chan<- struct{}) {
+func (w *lookupWalker) startNewLookup() {
+	if w.lookupDone != nil {
+		return // already running
+	}
+	for it := range w.iters {
+		if it.needsNodes() {
+			w.lookupDone = make(chan struct{})
+			go w.runLookup()
+			return
+		}
+	}
+	// all iterators have full buffer
+}
+
+func (w *lookupWalker) runLookup() {
 	w.lookup(func(n *enode.Node) {
 		select {
-		case nodes <- n:
+		case w.foundNode <- n:
 		case <-w.closeCh:
 		}
 	})
-	done <- struct{}{}
+	w.lookupDone <- struct{}{}
 }
 
 // lookupIterator is a sequence of discovered nodes.
 type lookupIterator struct {
-	w         *lookupWalker
-	buf       chan *enode.Node
-	closed    bool
-	closeOnce sync.Once
+	cur  *enode.Node
+	w    *lookupWalker
+	mu   sync.Mutex
+	cond *sync.Cond
+	buf  []*enode.Node
 }
 
 const lookupIteratorBuffer = 100
 
 func (w *lookupWalker) newIterator() *lookupIterator {
-	it := &lookupIterator{w: w, buf: make(chan *enode.Node, lookupIteratorBuffer)}
+	it := &lookupIterator{w: w, buf: make([]*enode.Node, 0, lookupIteratorBuffer)}
+	it.cond = sync.NewCond(&it.mu)
 	select {
 	case w.newIterCh <- it:
 	case <-w.closeCh:
-		it.closed = true
-		close(it.buf)
+		it.buf = nil
 	}
 	return it
 }
 
-func (it *lookupIterator) NextNode(ctx context.Context) (n *enode.Node, isLive bool) {
-	for {
-		select {
-		case it.w.triggerCh <- struct{}{}:
-			// lookup triggered
-		case n, ok := <-it.buf:
-			if !ok {
-				it.closed = true
-			}
-			return n, !it.closed
-		case <-ctx.Done():
-			return nil, !it.closed
-		}
+func (it *lookupIterator) Next() bool {
+	select {
+	case it.w.triggerCh <- struct{}{}:
+	case <-it.w.closeCh:
 	}
+	it.cur = nil
+
+	// Wait for the buffer to be filled.
+	it.mu.Lock()
+	defer it.mu.Unlock()
+	for it.buf != nil && len(it.buf) == 0 {
+		it.cond.Wait()
+	}
+	if it.buf == nil {
+		return false // closed
+	}
+	it.cur = it.buf[0]
+	copy(it.buf, it.buf[1:])
+	it.buf = it.buf[:len(it.buf)-1]
+	return true
+}
+
+func (it *lookupIterator) Node() *enode.Node {
+	return it.cur
 }
 
 func (it *lookupIterator) Close() {
-	it.closeOnce.Do(func() {
-		select {
-		case it.w.delIterCh <- it:
-		case <-it.w.closeCh:
-		}
-		it.drainAndClose()
-	})
+	select {
+	case it.w.delIterCh <- it:
+	case <-it.w.closeCh:
+	}
+	it.close()
+}
+
+func (it *lookupIterator) close() {
+	it.mu.Lock()
+	defer it.mu.Unlock()
+
+	if it.buf != nil {
+		it.buf = nil
+		it.cond.Signal()
+	}
 }
 
 // deliver sends a node to the iterator buffer.
 func (it *lookupIterator) deliver(n *enode.Node) {
-	// We don't want deliver to block and replace stale results when they're not being
-	// read. Check whether the buffer is full and allow one receive from the buffer if so.
-	// This is OK because there is only one writer.
-	var remove chan *enode.Node
-	if len(it.buf) == cap(it.buf) {
-		remove = it.buf
-	}
-	for {
-		select {
-		case it.buf <- n:
-			return
-		case <-remove:
-			remove = nil
-		}
+	it.mu.Lock()
+	defer it.mu.Unlock()
+
+	if len(it.buf) == lookupIteratorBuffer {
+		it.buf[rand.Intn(len(it.buf))] = n
+	} else {
+		it.buf = append(it.buf, n)
 	}
+	it.cond.Signal()
 }
 
-func (it *lookupIterator) drainAndClose() {
-	for len(it.buf) > 0 {
-		<-it.buf
-	}
-	close(it.buf)
+func (it *lookupIterator) needsNodes() bool {
+	it.mu.Lock()
+	defer it.mu.Unlock()
+
+	return len(it.buf) < lookupIteratorBuffer/3
 }
diff --git a/p2p/discover/common_test.go b/p2p/discover/common_test.go
index 3c55256d11..28df182755 100644
--- a/p2p/discover/common_test.go
+++ b/p2p/discover/common_test.go
@@ -17,7 +17,6 @@
 package discover
 
 import (
-	"context"
 	"encoding/binary"
 	"sync"
 	"sync/atomic"
@@ -33,32 +32,33 @@ import (
 func TestLookupIterator(t *testing.T) {
 	var (
 		test      = newLookupWalkerTest()
-		testNodes = make([]*enode.Node, 100)
+		testNodes = makeTestNodes(lookupIteratorBuffer)
 		wg        sync.WaitGroup
 	)
-	for i := range testNodes {
-		testNodes[i] = testNode(i)
-	}
+
 	testIterator := func(it discutil.Iterator) {
 		defer wg.Done()
 
-		ctx, cancel := context.WithTimeout(context.Background(), 2*time.Second)
-		defer cancel()
-		nodes := discutil.ReadNodes(ctx, it, 20)
-		sortByID(nodes) // ReadNodes may shuffle results
+		// Check reading nodes:
+		nodes := discutil.ReadNodes(it, 20)
+		sortByID(nodes)
 		if err := checkNodesEqual(nodes, testNodes[:20]); err != nil {
 			t.Error(err)
 		}
+		nodes = discutil.ReadNodes(it, 20)
+		sortByID(nodes)
+		if err := checkNodesEqual(nodes, testNodes[20:40]); err != nil {
+			t.Error(err)
+		}
 
+		// Check close:
 		it.Close()
-		n, isLive := it.NextNode(context.Background())
-		if n != nil {
-			t.Error("iterator returned non-nil node after close")
+		if it.Next() {
+			t.Error("Next returned true after close")
 		}
-		if isLive {
-			t.Error("iterator returned isLive == true after close")
+		if it.Node() != nil {
+			t.Error("iterator has non-nil node after close")
 		}
-
 		it.Close() // shouldn't crash
 	}
 
@@ -68,30 +68,35 @@ func TestLookupIterator(t *testing.T) {
 	}
 
 	test.serveOneLookup(testNodes[:10])
-	test.serveOneLookup(testNodes[10:])
+	test.serveOneLookup(testNodes[10:20])
+	test.serveOneLookup(testNodes[20:])
 	wg.Wait()
 
 	test.close()
 }
 
+func TestLookupIteratorClose(t *testing.T) {
+	test := newLookupWalkerTest()
+	defer test.close()
+	it := test.newIterator()
+
+	go func() {
+		time.Sleep(200 * time.Millisecond)
+		it.Close()
+	}()
+	it.Next()
+}
+
 // This test checks that the lookup iterator drops nodes when they're not being
 // read fast enough.
 func TestLookupIteratorDropStale(t *testing.T) {
 	var (
 		test       = newLookupWalkerTest()
-		testNodes  = make([]*enode.Node, 2*lookupIteratorBuffer)
+		testNodes  = makeTestNodes(2 * lookupIteratorBuffer)
 		lookupDone = make(chan struct{})
 	)
 	defer test.close()
-
-	for i := range testNodes {
-		testNodes[i] = testNode(i)
-	}
-
-	// Create iterator first so all found nodes go through its buffer.
 	it := test.newIterator()
-
-	// Serve one lookup.
 	go func() {
 		test.serveOneLookup(testNodes)
 		close(lookupDone)
@@ -99,17 +104,17 @@ func TestLookupIteratorDropStale(t *testing.T) {
 
 	// The first call to NextNode triggers the lookup and receives the first result
 	// as soon as it becomes available.
-	n, _ := it.NextNode(context.Background())
-	if n != testNodes[0] {
-		t.Fatalf("wrong result %d: got %v, want %v", 0, n.ID(), testNodes[0])
+	it.Next()
+	if it.Node() != testNodes[0] {
+		t.Fatalf("wrong result %d: got %v, want %v", 0, it.Node().ID(), testNodes[0])
 	}
 
 	// Now wait for the lookup to finish and read the remaining nodes.
 	<-lookupDone
 	for i := 0; i < lookupIteratorBuffer; i++ {
-		n, _ := it.NextNode(context.Background())
+		it.Next()
 		for _, tn := range testNodes[lookupIteratorBuffer:] {
-			if n == tn {
+			if it.Node() == tn {
 				return
 			}
 		}
@@ -117,10 +122,40 @@ func TestLookupIteratorDropStale(t *testing.T) {
 	t.Fatal("didn't find any node from second half of testNodes")
 }
 
-func testNode(id int) *enode.Node {
-	var nodeID enode.ID
-	binary.BigEndian.PutUint64(nodeID[:], uint64(id))
-	return enode.SignNull(new(enr.Record), nodeID)
+// This test checks that the iterator kicks off a lookup when Next is called.
+func TestLookupIteratorDrained(t *testing.T) {
+	var (
+		test      = newLookupWalkerTest()
+		it        = test.newIterator()
+		testNodes = makeTestNodes(2 * lookupIteratorBuffer)
+	)
+
+	test.serveOneLookup(testNodes[:lookupIteratorBuffer])
+	nodes := discutil.ReadNodes(it, lookupIteratorBuffer)
+	sortByID(nodes)
+	if err := checkNodesEqual(nodes, testNodes[:lookupIteratorBuffer]); err != nil {
+		t.Fatal(err)
+	}
+
+	// Here the iterator buffer is drained and no lookup is running.
+
+	// Request more nodes. This needs to start another lookup.
+	go test.serveOneLookup(testNodes[lookupIteratorBuffer:])
+	nodes = discutil.ReadNodes(it, 10)
+	sortByID(nodes)
+	if err := checkNodesEqual(nodes, testNodes[lookupIteratorBuffer:lookupIteratorBuffer+10]); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func makeTestNodes(n int) []*enode.Node {
+	nodes := make([]*enode.Node, n)
+	for i := range nodes {
+		var nodeID enode.ID
+		binary.BigEndian.PutUint64(nodeID[:], uint64(i))
+		nodes[i] = enode.SignNull(new(enr.Record), nodeID)
+	}
+	return nodes
 }
 
 type lookupWalkerTest struct {