Merge pull request #51925 from fuweid/fix-46212

daemon: ignore duplicate task exit events in daemon state

daemon/container/state.go

@@ -299,7 +299,11 @@ func (s *State) SetRestarting(exitStatus *ExitStatus) {
 	s.Restarting = true
 	s.Paused = false
 	s.Pid = 0
-	s.FinishedAt = time.Now().UTC()
+	finishedAt := exitStatus.ExitedAt
+	if finishedAt.IsZero() {
+		finishedAt = time.Now().UTC()
+	}
+	s.FinishedAt = finishedAt
 	s.ExitCode = exitStatus.ExitCode
 
 	s.notifyAndClear(&s.stopWaiters)

daemon/monitor.go

@@ -48,6 +48,13 @@ func (daemon *Daemon) handleContainerExit(c *container.Container, e *libcontaine
 		return nil
 	}
 
+	// Ignore duplicate exit event that may arrive after the first one.
+	// See moby/moby#46212.
+	if daemon.shouldIgnoreExitEventWithLock(c, e) {
+		c.Unlock()
+		return nil
+	}
+
 	cfg := daemon.config()
 
 	// Health checks will be automatically restarted if/when the
@@ -339,3 +346,50 @@ func (daemon *Daemon) autoRemove(cfg *config.Config, c *container.Container) {
 		log.G(context.TODO()).WithFields(log.Fields{"error": err, "container": c.ID}).Error("error removing container")
 	}
 }
+
+func (daemon *Daemon) shouldIgnoreExitEventWithLock(c *container.Container, e *libcontainerdtypes.EventInfo) (ret bool) {
+	if e == nil {
+		return false
+	}
+
+	curState := c.State.State()
+
+	defer func() {
+		if ret {
+			log.G(context.TODO()).
+				WithFields(log.Fields{
+					"container": c.ID,
+					"state":     c.State.String(),
+					"exitCode":  e.ExitCode,
+					"exitedAt":  e.ExitedAt,
+				}).Info("ignoring duplicate container exit event")
+		}
+	}()
+
+	switch curState {
+	case containertypes.StateRemoving,
+		containertypes.StateExited,
+		containertypes.StateDead:
+
+		return true
+
+	case containertypes.StateRunning:
+		// If the container is running, but the exit event is from
+		// before it was started, ignore it. This can happen when a
+		// duplicate exit arrives while the restart path holds the
+		// container lock; by the time we process it, a new task is
+		// already running, so the exit belongs to the previous task.
+		return !e.ExitedAt.IsZero() && e.ExitedAt.Before(c.StartedAt)
+
+	case containertypes.StateRestarting:
+		// The restart path acquires and holds the container lock before
+		// processing; on failure it transitions the container to exited,
+		// and on success it transitions to running. Therefore, any exit
+		// event observed while still restarting is a late duplicate from
+		// the previous task and should be ignored.
+		return !e.ExitedAt.IsZero() && e.ExitedAt.After(c.FinishedAt)
+
+	default:
+		return false
+	}
+}