control/controlclient: move watchdog out of mapSession

In prep for making mapSession's lifetime not be 1:1 with a single HTTP
response's lifetime, this moves the inactivity timer watchdog out of
mapSession and into the caller that owns the streaming HTTP response.

(This is admittedly closer to how it was prior to the mapSession type
existing, but that was before we connected some dots which were
impossible to even see before the mapSession type broke the code up.)

Updates #7175

Change-Id: Ia108dac84a4953db41cbd30e73b1de4a2a676c11
Signed-off-by: Brad Fitzpatrick <bradfitz@tailscale.com>

control/controlclient/direct.go

@@ -828,7 +828,7 @@ const watchdogTimeout = 120 * time.Second
 // if the context expires or the server returns an error/closes the connection
 // and as such always returns a non-nil error.
 //
-// If cb is nil, OmitPeers will be set to true.
+// If nu is nil, OmitPeers will be set to true.
 func (c *Direct) sendMapRequest(ctx context.Context, isStreaming bool, nu NetmapUpdater) error {
 	if isStreaming && nu == nil {
 		panic("cb must be non-nil if isStreaming is true")
@@ -992,7 +992,27 @@ func (c *Direct) sendMapRequest(ctx context.Context, isStreaming bool, nu Netmap
 		}
 		c.expiry = nm.Expiry
 	}
-	sess.StartWatchdog()
+
+	// Create a watchdog timer that breaks the connection if we don't receive a
+	// MapResponse from the network at least once every two minutes. The
+	// watchdog timer is stopped every time we receive a MapResponse (so it
+	// doesn't run when we're processing a MapResponse message, including any
+	// long-running requested operations like Debug.Sleep) and is reset whenever
+	// we go back to blocking on network reads.
+	watchdogTimer, watchdogTimedOut := c.clock.NewTimer(watchdogTimeout)
+	defer watchdogTimer.Stop()
+
+	go func() {
+		select {
+		case <-ctx.Done():
+			vlogf("netmap: ending timeout goroutine")
+			return
+		case <-watchdogTimedOut:
+			c.logf("map response long-poll timed out!")
+			cancel()
+			return
+		}
+	}()
 
 	// gotNonKeepAliveMessage is whether we've yet received a MapResponse message without
 	// KeepAlive set.
@@ -1006,6 +1026,7 @@ func (c *Direct) sendMapRequest(ctx context.Context, isStreaming bool, nu Netmap
 	// We can use this same read loop either way.
 	var msg []byte
 	for mapResIdx := 0; mapResIdx == 0 || isStreaming; mapResIdx++ {
+		watchdogTimer.Reset(watchdogTimeout)
 		vlogf("netmap: starting size read after %v (poll %v)", time.Since(t0).Round(time.Millisecond), mapResIdx)
 		var siz [4]byte
 		if _, err := io.ReadFull(res.Body, siz[:]); err != nil {
@@ -1026,6 +1047,7 @@ func (c *Direct) sendMapRequest(ctx context.Context, isStreaming bool, nu Netmap
 			vlogf("netmap: decode error: %v")
 			return err
 		}
+		watchdogTimer.Stop()
 
 		metricMapResponseMessages.Add(1)
 
@@ -1068,14 +1090,6 @@ func (c *Direct) sendMapRequest(ctx context.Context, isStreaming bool, nu Netmap
 				c.logf("netmap: [unexpected] new dial plan; nowhere to store it")
 			}
 		}
-
-		select {
-		case sess.watchdogReset <- struct{}{}:
-			vlogf("netmap: sent timer reset")
-		case <-ctx.Done():
-			c.logf("[v1] netmap: not resetting timer; context done: %v", ctx.Err())
-			return ctx.Err()
-		}
 		if resp.KeepAlive {
 			metricMapResponseKeepAlives.Add(1)
 			continue
@@ -1102,7 +1116,7 @@ func (c *Direct) sendMapRequest(ctx context.Context, isStreaming bool, nu Netmap
 	return nil
 }
 
-func (c *Direct) handleDebugMessage(ctx context.Context, debug *tailcfg.Debug, watchdogReset chan<- struct{}) error {
+func (c *Direct) handleDebugMessage(ctx context.Context, debug *tailcfg.Debug) error {
 	if code := debug.Exit; code != nil {
 		c.logf("exiting process with status %v per controlplane", *code)
 		os.Exit(*code)
@@ -1112,7 +1126,7 @@ func (c *Direct) handleDebugMessage(ctx context.Context, debug *tailcfg.Debug, w
 		envknob.SetNoLogsNoSupport()
 	}
 	if sleep := time.Duration(debug.SleepSeconds * float64(time.Second)); sleep > 0 {
-		if err := sleepAsRequested(ctx, c.logf, watchdogReset, sleep, c.clock); err != nil {
+		if err := sleepAsRequested(ctx, c.logf, sleep, c.clock); err != nil {
 			return err
 		}
 	}
@@ -1444,7 +1458,7 @@ func answerC2NPing(logf logger.Logf, c2nHandler http.Handler, c *http.Client, pr
 // that the client sleep. The complication is that while we're sleeping (if for
 // a long time), we need to periodically reset the watchdog timer before it
 // expires.
-func sleepAsRequested(ctx context.Context, logf logger.Logf, watchdogReset chan<- struct{}, d time.Duration, clock tstime.Clock) error {
+func sleepAsRequested(ctx context.Context, logf logger.Logf, d time.Duration, clock tstime.Clock) error {
 	const maxSleep = 5 * time.Minute
 	if d > maxSleep {
 		logf("sleeping for %v, capped from server-requested %v ...", maxSleep, d)
@@ -1453,25 +1467,13 @@ func sleepAsRequested(ctx context.Context, logf logger.Logf, watchdogReset chan<
 		logf("sleeping for server-requested %v ...", d)
 	}
 
-	ticker, tickerChannel := clock.NewTicker(watchdogTimeout / 2)
-	defer ticker.Stop()
 	timer, timerChannel := clock.NewTimer(d)
 	defer timer.Stop()
-	for {
-		select {
-		case <-ctx.Done():
-			return ctx.Err()
-		case <-timerChannel:
-			return nil
-		case <-tickerChannel:
-			select {
-			case watchdogReset <- struct{}{}:
-			case <-timerChannel:
-				return nil
-			case <-ctx.Done():
-				return ctx.Err()
-			}
-		}
+	select {
+	case <-ctx.Done():
+		return ctx.Err()
+	case <-timerChannel:
+		return nil
 	}
 }