Implement proper shutdown logic

Signed-off-by: Michael Crosby <crosbymichael@gmail.com>

api/v1/server.go

@@ -325,10 +325,14 @@ func (s *server) createCheckpoint(w http.ResponseWriter, r *http.Request) {
 		name = vars["name"]
 	)
 	var cp Checkpoint
+	// most options to the checkpoint action can be left out so don't
+	// decode unless the client passed anything in the body.
+	if r.ContentLength > 0 {
 		if err := json.NewDecoder(r.Body).Decode(&cp); err != nil {
 			http.Error(w, err.Error(), http.StatusBadRequest)
 			return
 		}
+	}
 	e := containerd.NewEvent(containerd.CreateCheckpointEventType)
 	e.ID = id
 	e.Checkpoint = &runtime.Checkpoint{

containerd/main.go

@@ -86,9 +86,12 @@ func daemon(stateDir string, concurrency, bufferSize int) error {
 		w := containerd.NewWorker(supervisor, wg)
 		go w.Start()
 	}
+	// only set containerd as the subreaper if it is not an init process
+	if pid := os.Getpid(); pid != 1 {
 		if err := setSubReaper(); err != nil {
 			return err
 		}
+	}
 	// start the signal handler in the background.
 	go startSignalHandler(supervisor, bufferSize)
 	if err := supervisor.Start(); err != nil {

containerd/reap_linux.go

@@ -34,8 +34,7 @@ func startSignalHandler(supervisor *containerd.Supervisor, bufferSize int) {
 	for s := range signals {
 		switch s {
 		case syscall.SIGTERM, syscall.SIGINT:
-			supervisor.Close()
+			supervisor.Stop(signals)
-			os.Exit(0)
 		case syscall.SIGCHLD:
 			exits, err := reap()
 			if err != nil {
@@ -46,6 +45,8 @@ func startSignalHandler(supervisor *containerd.Supervisor, bufferSize int) {
 			}
 		}
 	}
+	supervisor.Close()
+	os.Exit(0)
 }
 
 func reap() (exits []*containerd.Event, err error) {

delete.go

@@ -15,6 +15,7 @@ func (h *DeleteEvent) Handle(e *Event) error {
 			logrus.WithField("error", err).Error("containerd: deleting container")
 		}
 		ContainersCounter.Dec(1)
+		h.s.containerGroup.Done()
 	}
 	return nil
 }

supervisor.go

@@ -2,8 +2,11 @@ package containerd
 
 import (
 	"os"
+	"os/signal"
 	"path/filepath"
 	goruntime "runtime"
+	"sync"
+	"syscall"
 
 	"github.com/Sirupsen/logrus"
 	"github.com/docker/containerd/runtime"
@@ -66,14 +69,48 @@ type Supervisor struct {
 	tasks          chan *StartTask
 	subscribers    map[subscriber]bool
 	machine        Machine
+	containerGroup sync.WaitGroup
 }
 
 type subscriber chan *Event
 
-// need proper close logic for jobs and stuff so that sending to the channels dont panic
+func (s *Supervisor) Stop(sig chan os.Signal) {
-// but can complete jobs
+	// Close the tasks channel so that no new containers get started
+	close(s.tasks)
+	// send a SIGTERM to all containers
+	for id, c := range s.containers {
+		logrus.WithField("id", id).Debug("sending TERM to container processes")
+		procs, err := c.Processes()
+		if err != nil {
+			logrus.WithField("id", id).Warn("get container processes")
+			continue
+		}
+		if len(procs) == 0 {
+			continue
+		}
+		mainProc := procs[0]
+		if err := mainProc.Signal(syscall.SIGTERM); err != nil {
+			pid, _ := mainProc.Pid()
+			logrus.WithFields(logrus.Fields{
+				"id":    id,
+				"pid":   pid,
+				"error": err,
+			}).Error("send SIGTERM to process")
+		}
+	}
+	go func() {
+		logrus.Debug("waiting for containers to exit")
+		s.containerGroup.Wait()
+		logrus.Debug("all containers exited")
+		// stop receiving signals and close the channel
+		signal.Stop(sig)
+		close(sig)
+	}()
+}
+
+// Close closes any open files in the supervisor but expects that Stop has been
+// callsed so that no more containers are started.
 func (s *Supervisor) Close() error {
-	//TODO: unsubscribe all channels
 	return s.journal.Close()
 }
 

worker.go

@@ -55,6 +55,7 @@ func (w *worker) Start() {
 				continue
 			}
 		}
+		w.s.containerGroup.Add(1)
 		ContainerStartTimer.UpdateSince(started)
 		t.Err <- nil
 	}