device: use channel close to shut down and drain encryption channel

The new test introduced in this commit used to deadlock about 1% of the time.

I believe that the deadlock occurs as follows:

* The test completes, calling device.Close.
* device.Close closes device.signals.stop.
* RoutineEncryption stops.
* The deferred function in RoutineEncryption drains device.queue.encryption.
* RoutineEncryption exits.
* A peer's RoutineNonce processes an element queued in peer.queue.nonce.
* RoutineNonce puts that element into the outbound and encryption queues.
* RoutineSequentialSender reads that elements from the outbound queue.
* It waits for that element to get Unlocked by RoutineEncryption.
* RoutineEncryption has already exited, so RoutineSequentialSender blocks forever.
* device.RemoveAllPeers calls peer.Stop on all peers.
* peer.Stop waits for peer.routines.stopping, which blocks forever.

Rather than attempt to add even more ordering to the already complex
centralized shutdown orchestration, this commit moves towards a
data-flow-oriented shutdown.

The device.queue.encryption gets closed when there will be no more writes to it.
All device.queue.encryption readers always read until the channel is closed and then exit.
We thus guarantee that any element that enters the encryption queue also exits it.
This removes the need for central control of the lifetime of RoutineEncryption,
removes the need to drain the encryption queue on shutdown, and simplifies RoutineEncryption.

This commit also fixes a data race. When RoutineSequentialSender
drains its queue on shutdown, it needs to lock the elem before operating on it,
just as the main body does.

The new test in this commit passed 50k iterations with the race detector enabled
and 150k iterations with the race detector disabled, with no failures.

Signed-off-by: Josh Bleecher Snyder <josh@tailscale.com>

3 files changed, 170 insertions, 98 deletions

diff --git a/device/device.go b/device/device.go
index 9e2d001..d9367e5 100644
--- a/device/device.go
+++ b/device/device.go
@@ -74,7 +74,7 @@ type Device struct {
 	}
 
 	queue struct {
-		encryption chan *QueueOutboundElement
+		encryption *encryptionQueue
 		decryption chan *QueueInboundElement
 		handshake  chan QueueHandshakeElement
 	}
@@ -89,6 +89,31 @@ type Device struct {
 	}
 }
 
+// An encryptionQueue is a channel of QueueOutboundElements awaiting encryption.
+// An encryptionQueue is ref-counted using its wg field.
+// An encryptionQueue created with newEncryptionQueue has one reference.
+// Every additional writer must call wg.Add(1).
+// Every completed writer must call wg.Done().
+// When no further writers will be added,
+// call wg.Done to remove the initial reference.
+// When the refcount hits 0, the queue's channel is closed.
+type encryptionQueue struct {
+	c  chan *QueueOutboundElement
+	wg sync.WaitGroup
+}
+
+func newEncryptionQueue() *encryptionQueue {
+	q := &encryptionQueue{
+		c: make(chan *QueueOutboundElement, QueueOutboundSize),
+	}
+	q.wg.Add(1)
+	go func() {
+		q.wg.Wait()
+		close(q.c)
+	}()
+	return q
+}
+
 /* Converts the peer into a "zombie", which remains in the peer map,
  * but processes no packets and does not exists in the routing table.
  *
@@ -280,7 +305,7 @@ func NewDevice(tunDevice tun.Device, logger *Logger) *Device {
 	// create queues
 
 	device.queue.handshake = make(chan QueueHandshakeElement, QueueHandshakeSize)
-	device.queue.encryption = make(chan *QueueOutboundElement, QueueOutboundSize)
+	device.queue.encryption = newEncryptionQueue()
 	device.queue.decryption = make(chan *QueueInboundElement, QueueInboundSize)
 
 	// prepare signals
@@ -297,7 +322,7 @@ func NewDevice(tunDevice tun.Device, logger *Logger) *Device {
 	cpus := runtime.NumCPU()
 	device.state.stopping.Wait()
 	for i := 0; i < cpus; i += 1 {
-		device.state.stopping.Add(3)
+		device.state.stopping.Add(2) // decryption and handshake
 		go device.RoutineEncryption()
 		go device.RoutineDecryption()
 		go device.RoutineHandshake()
@@ -346,10 +371,6 @@ func (device *Device) FlushPacketQueues() {
 			if ok {
 				elem.Drop()
 			}
-		case elem, ok := <-device.queue.encryption:
-			if ok {
-				elem.Drop()
-			}
 		case <-device.queue.handshake:
 		default:
 			return
@@ -373,6 +394,10 @@ func (device *Device) Close() {
 
 	device.isUp.Set(false)
 
+	// We kept a reference to the encryption queue,
+	// in case we started any new peers that might write to it.
+	// No new peers are coming; we are done with the encryption queue.
+	device.queue.encryption.wg.Done()
 	close(device.signals.stop)
 	device.state.stopping.Wait()
 
diff --git a/device/device_test.go b/device/device_test.go
index a89dcc2..65942ec 100644
--- a/device/device_test.go
+++ b/device/device_test.go
@@ -7,9 +7,11 @@ package device
 
 import (
 	"bytes"
+	"errors"
 	"fmt"
 	"io"
 	"net"
+	"sync"
 	"testing"
 	"time"
 
@@ -79,18 +81,74 @@ func genConfigs(t *testing.T) (cfgs [2]io.Reader) {
 	return
 }
 
-// genChannelTUNs creates a usable pair of ChannelTUNs for use in a test.
-func genChannelTUNs(t *testing.T) (tun [2]*tuntest.ChannelTUN) {
+// A testPair is a pair of testPeers.
+type testPair [2]testPeer
+
+// A testPeer is a peer used for testing.
+type testPeer struct {
+	tun *tuntest.ChannelTUN
+	dev *Device
+	ip  net.IP
+}
+
+type SendDirection bool
+
+const (
+	Ping SendDirection = true
+	Pong SendDirection = false
+)
+
+func (pair *testPair) Send(t *testing.T, ping SendDirection, done chan struct{}) {
+	t.Helper()
+	p0, p1 := pair[0], pair[1]
+	if !ping {
+		// pong is the new ping
+		p0, p1 = p1, p0
+	}
+	msg := tuntest.Ping(p0.ip, p1.ip)
+	p1.tun.Outbound <- msg
+	timer := time.NewTimer(5 * time.Second)
+	defer timer.Stop()
+	var err error
+	select {
+	case msgRecv := <-p0.tun.Inbound:
+		if !bytes.Equal(msg, msgRecv) {
+			err = errors.New("ping did not transit correctly")
+		}
+	case <-timer.C:
+		err = errors.New("ping did not transit")
+	case <-done:
+	}
+	if err != nil {
+		// The error may have occurred because the test is done.
+		select {
+		case <-done:
+			return
+		default:
+		}
+		// Real error.
+		t.Error(err)
+	}
+}
+
+// genTestPair creates a testPair.
+func genTestPair(t *testing.T) (pair testPair) {
 	const maxAttempts = 10
 NextAttempt:
 	for i := 0; i < maxAttempts; i++ {
 		cfg := genConfigs(t)
 		// Bring up a ChannelTun for each config.
-		for i := range tun {
-			tun[i] = tuntest.NewChannelTUN()
-			dev := NewDevice(tun[i].TUN(), NewLogger(LogLevelDebug, fmt.Sprintf("dev%d: ", i)))
-			dev.Up()
-			if err := dev.IpcSetOperation(cfg[i]); err != nil {
+		for i := range pair {
+			p := &pair[i]
+			p.tun = tuntest.NewChannelTUN()
+			if i == 0 {
+				p.ip = net.ParseIP("1.0.0.1")
+			} else {
+				p.ip = net.ParseIP("1.0.0.2")
+			}
+			p.dev = NewDevice(p.tun.TUN(), NewLogger(LogLevelDebug, fmt.Sprintf("dev%d: ", i)))
+			p.dev.Up()
+			if err := p.dev.IpcSetOperation(cfg[i]); err != nil {
 				// genConfigs attempted to pick ports that were free.
 				// There's a tiny window between genConfigs closing the port
 				// and us opening it, during which another process could
@@ -104,12 +162,12 @@ NextAttempt:
 			// The device might still not be up, e.g. due to an error
 			// in RoutineTUNEventReader's call to dev.Up that got swallowed.
 			// Assume it's due to a transient error (port in use), and retry.
-			if !dev.isUp.Get() {
-				t.Logf("%v did not come up, trying again", dev)
+			if !p.dev.isUp.Get() {
+				t.Logf("device %d did not come up, trying again", i)
 				continue NextAttempt
 			}
 			// The device is up. Close it when the test completes.
-			t.Cleanup(dev.Close)
+			t.Cleanup(p.dev.Close)
 		}
 		return // success
 	}
@@ -119,33 +177,47 @@ NextAttempt:
 }
 
 func TestTwoDevicePing(t *testing.T) {
-	tun := genChannelTUNs(t)
-
+	pair := genTestPair(t)
 	t.Run("ping 1.0.0.1", func(t *testing.T) {
-		msg2to1 := tuntest.Ping(net.ParseIP("1.0.0.1"), net.ParseIP("1.0.0.2"))
-		tun[1].Outbound <- msg2to1
-		select {
-		case msgRecv := <-tun[0].Inbound:
-			if !bytes.Equal(msg2to1, msgRecv) {
-				t.Error("ping did not transit correctly")
-			}
-		case <-time.After(5 * time.Second):
-			t.Error("ping did not transit")
-		}
+		pair.Send(t, Ping, nil)
 	})
-
 	t.Run("ping 1.0.0.2", func(t *testing.T) {
-		msg1to2 := tuntest.Ping(net.ParseIP("1.0.0.2"), net.ParseIP("1.0.0.1"))
-		tun[0].Outbound <- msg1to2
-		select {
-		case msgRecv := <-tun[1].Inbound:
-			if !bytes.Equal(msg1to2, msgRecv) {
-				t.Error("return ping did not transit correctly")
+		pair.Send(t, Pong, nil)
+	})
+}
+
+// TestConcurrencySafety does other things concurrently with tunnel use.
+// It is intended to be used with the race detector to catch data races.
+func TestConcurrencySafety(t *testing.T) {
+	pair := genTestPair(t)
+	done := make(chan struct{})
+
+	const warmupIters = 10
+	var warmup sync.WaitGroup
+	warmup.Add(warmupIters)
+	go func() {
+		// Send data continuously back and forth until we're done.
+		// Note that we may continue to attempt to send data
+		// even after done is closed.
+		i := warmupIters
+		for ping := Ping; ; ping = !ping {
+			pair.Send(t, ping, done)
+			select {
+			case <-done:
+				return
+			default:
+			}
+			if i > 0 {
+				warmup.Done()
+				i--
 			}
-		case <-time.After(5 * time.Second):
-			t.Error("return ping did not transit")
 		}
-	})
+	}()
+	warmup.Wait()
+
+	// coming soon: more things here...
+
+	close(done)
 }
 
 func assertNil(t *testing.T, err error) {
diff --git a/device/send.go b/device/send.go
index 0801b71..1b16edd 100644
--- a/device/send.go
+++ b/device/send.go
@@ -352,6 +352,9 @@ func (peer *Peer) RoutineNonce() {
 	device := peer.device
 	logDebug := device.log.Debug
 
+	// We write to the encryption queue; keep it alive until we are done.
+	device.queue.encryption.wg.Add(1)
+
 	flush := func() {
 		for {
 			select {
@@ -368,6 +371,7 @@ func (peer *Peer) RoutineNonce() {
 		flush()
 		logDebug.Println(peer, "- Routine: nonce worker - stopped")
 		peer.queue.packetInNonceQueueIsAwaitingKey.Set(false)
+		device.queue.encryption.wg.Done() // no more writes from us
 		peer.routines.stopping.Done()
 	}()
 
@@ -455,7 +459,7 @@ NextPacket:
 			elem.Lock()
 
 			// add to parallel and sequential queue
-			addToOutboundAndEncryptionQueues(peer.queue.outbound, device.queue.encryption, elem)
+			addToOutboundAndEncryptionQueues(peer.queue.outbound, device.queue.encryption.c, elem)
 		}
 	}
 }
@@ -486,76 +490,46 @@ func (device *Device) RoutineEncryption() {
 
 	logDebug := device.log.Debug
 
-	defer func() {
-		for {
-			select {
-			case elem, ok := <-device.queue.encryption:
-				if ok && !elem.IsDropped() {
-					elem.Drop()
-					device.PutMessageBuffer(elem.buffer)
-					elem.Unlock()
-				}
-			default:
-				goto out
-			}
-		}
-	out:
-		logDebug.Println("Routine: encryption worker - stopped")
-		device.state.stopping.Done()
-	}()
-
+	defer logDebug.Println("Routine: encryption worker - stopped")
 	logDebug.Println("Routine: encryption worker - started")
 
-	for {
-
-		// fetch next element
+	for elem := range device.queue.encryption.c {
 
-		select {
-		case <-device.signals.stop:
-			return
-
-		case elem, ok := <-device.queue.encryption:
-
-			if !ok {
-				return
-			}
-
-			// check if dropped
+		// check if dropped
 
-			if elem.IsDropped() {
-				continue
-			}
+		if elem.IsDropped() {
+			continue
+		}
 
-			// populate header fields
+		// populate header fields
 
-			header := elem.buffer[:MessageTransportHeaderSize]
+		header := elem.buffer[:MessageTransportHeaderSize]
 
-			fieldType := header[0:4]
-			fieldReceiver := header[4:8]
-			fieldNonce := header[8:16]
+		fieldType := header[0:4]
+		fieldReceiver := header[4:8]
+		fieldNonce := header[8:16]
 
-			binary.LittleEndian.PutUint32(fieldType, MessageTransportType)
-			binary.LittleEndian.PutUint32(fieldReceiver, elem.keypair.remoteIndex)
-			binary.LittleEndian.PutUint64(fieldNonce, elem.nonce)
+		binary.LittleEndian.PutUint32(fieldType, MessageTransportType)
+		binary.LittleEndian.PutUint32(fieldReceiver, elem.keypair.remoteIndex)
+		binary.LittleEndian.PutUint64(fieldNonce, elem.nonce)
 
-			// pad content to multiple of 16
+		// pad content to multiple of 16
 
-			paddingSize := calculatePaddingSize(len(elem.packet), int(atomic.LoadInt32(&device.tun.mtu)))
-			for i := 0; i < paddingSize; i++ {
-				elem.packet = append(elem.packet, 0)
-			}
+		paddingSize := calculatePaddingSize(len(elem.packet), int(atomic.LoadInt32(&device.tun.mtu)))
+		for i := 0; i < paddingSize; i++ {
+			elem.packet = append(elem.packet, 0)
+		}
 
-			// encrypt content and release to consumer
+		// encrypt content and release to consumer
 
-			binary.LittleEndian.PutUint64(nonce[4:], elem.nonce)
-			elem.packet = elem.keypair.send.Seal(
-				header,
-				nonce[:],
-				elem.packet,
-				nil,
-			)
-			elem.Unlock()
-		}
+		binary.LittleEndian.PutUint64(nonce[4:], elem.nonce)
+		elem.packet = elem.keypair.send.Seal(
+			header,
+			nonce[:],
+			elem.packet,
+			nil,
+		)
+		elem.Unlock()
 	}
 }
 
@@ -576,6 +550,7 @@ func (peer *Peer) RoutineSequentialSender() {
 			select {
 			case elem, ok := <-peer.queue.outbound:
 				if ok {
+					elem.Lock()
 					if !elem.IsDropped() {
 						device.PutMessageBuffer(elem.buffer)
 						elem.Drop()