cmd/syncthing, lib/events, lib/sync: Add timeout to REST event API, remove Ping (fixes #3933)

GitHub-Pull-Request: https://github.com/syncthing/syncthing/pull/3941

lib/sync/sync.go

@@ -227,3 +227,70 @@ func goid() int {
 	}
 	return id
 }
+
+// TimeoutCond is a variant on Cond. It has roughly the same semantics regarding 'L' - it must be held
+// both when broadcasting and when calling TimeoutCondWaiter.Wait()
+// Call Broadcast() to broadcast to all waiters on the TimeoutCond. Call SetupWait to create a
+// TimeoutCondWaiter configured with the given timeout, which can then be used to listen for
+// broadcasts.
+type TimeoutCond struct {
+	L  sync.Locker
+	ch chan struct{}
+}
+
+// TimeoutCondWaiter is a type allowing a consumer to wait on a TimeoutCond with a timeout. Wait() may be called multiple times,
+// and will return true every time that the TimeoutCond is broadcast to. Once the configured timeout
+// expires, Wait() will return false.
+// Call Stop() to release resources once this TimeoutCondWaiter is no longer needed.
+type TimeoutCondWaiter struct {
+	c     *TimeoutCond
+	timer *time.Timer
+}
+
+func NewTimeoutCond(l sync.Locker) *TimeoutCond {
+	return &TimeoutCond{
+		L: l,
+	}
+}
+
+func (c *TimeoutCond) Broadcast() {
+	// ch.L must be locked when calling this function
+
+	if c.ch != nil {
+		close(c.ch)
+		c.ch = nil
+	}
+}
+
+func (c *TimeoutCond) SetupWait(timeout time.Duration) *TimeoutCondWaiter {
+	timer := time.NewTimer(timeout)
+
+	return &TimeoutCondWaiter{
+		c:     c,
+		timer: timer,
+	}
+}
+
+func (w *TimeoutCondWaiter) Wait() bool {
+	// ch.L must be locked when calling this function
+
+	// Ensure that the channel exists, since we're going to be waiting on it
+	if w.c.ch == nil {
+		w.c.ch = make(chan struct{})
+	}
+	ch := w.c.ch
+
+	w.c.L.Unlock()
+	defer w.c.L.Lock()
+
+	select {
+	case <-w.timer.C:
+		return false
+	case <-ch:
+		return true
+	}
+}
+
+func (w *TimeoutCondWaiter) Stop() {
+	w.timer.Stop()
+}

lib/sync/sync_test.go

@@ -226,3 +226,100 @@ func TestWaitGroup(t *testing.T) {
 	debug = false
 	l.SetDebug("sync", false)
 }
+
+func TestTimeoutCond(t *testing.T) {
+	// WARNING this test relies heavily on threads not being stalled at particular points.
+	// As such, it's pretty unstable on the build server. It has been left in as it still
+	// exercises the deadlock detector, and one of the two things it tests is still functional.
+	// See the comments in runLocks
+
+	const (
+		// Low values to avoid being intrusive in continous testing. Can be
+		// increased significantly for stress testing.
+		iterations = 100
+		routines   = 10
+
+		timeMult = 2
+	)
+
+	c := NewTimeoutCond(NewMutex())
+
+	// Start a routine to periodically broadcast on the cond.
+
+	go func() {
+		d := time.Duration(routines) * timeMult * time.Millisecond / 2
+		t.Log("Broadcasting every", d)
+		for i := 0; i < iterations; i++ {
+			time.Sleep(d)
+
+			c.L.Lock()
+			c.Broadcast()
+			c.L.Unlock()
+		}
+	}()
+
+	// Start several routines that wait on it with different timeouts.
+
+	var results [routines][2]int
+	var wg sync.WaitGroup
+	for i := 0; i < routines; i++ {
+		i := i
+		wg.Add(1)
+		go func() {
+			d := time.Duration(i) * timeMult * time.Millisecond
+			t.Logf("Routine %d waits for %v\n", i, d)
+			succ, fail := runLocks(t, iterations, c, d)
+			results[i][0] = succ
+			results[i][1] = fail
+			wg.Done()
+		}()
+	}
+
+	wg.Wait()
+
+	// Print a table of routine number: successes, failures.
+
+	for i, v := range results {
+		t.Logf("%4d: %4d %4d\n", i, v[0], v[1])
+	}
+}
+
+func runLocks(t *testing.T, iterations int, c *TimeoutCond, d time.Duration) (succ, fail int) {
+	for i := 0; i < iterations; i++ {
+		c.L.Lock()
+
+		// The thread may be stalled, so we can't test the 'succeeded late' case reliably.
+		// Therefore make sure that we start t0 before starting the timeout, and only test
+		// the 'failed early' case.
+
+		t0 := time.Now()
+		w := c.SetupWait(d)
+
+		res := w.Wait()
+		waited := time.Since(t0)
+
+		// Allow 20% slide in either direction, and a five milliseconds of
+		// scheduling delay... In tweaking these it was clear that things
+		// worked like the should, so if this becomes a spurious failure
+		// kind of thing feel free to remove or give significantly more
+		// slack.
+
+		if !res && waited < d*8/10 {
+			t.Errorf("Wait failed early, %v < %v", waited, d)
+		}
+		if res && waited > d*11/10+5*time.Millisecond {
+			// Ideally this would be t.Errorf
+			t.Logf("WARNING: Wait succeeded late, %v > %v. This is probably a thread scheduling issue", waited, d)
+		}
+
+		w.Stop()
+
+		if res {
+			succ++
+		} else {
+			fail++
+		}
+		c.L.Unlock()
+	}
+	return
+}