Merge pull request #22947 from tonistiigi/fix-already-active-1.11.2

[v1.11.2] Fix already active error on restoring from ungraceful restart

integration-cli/docker_cli_daemon_not_experimental_test.go

@@ -37,3 +37,23 @@ func (s *DockerDaemonSuite) TestCleanupMountsAfterDaemonKill(c *check.C) {
 	comment = check.Commentf("%s is still mounted from older daemon start:\nDaemon root repository %s\n%s", id, s.d.folder, mountOut)
 	c.Assert(strings.Contains(string(mountOut), id), check.Equals, false, comment)
 }
+
+// #22913
+func (s *DockerDaemonSuite) TestContainerStartAfterDaemonKill(c *check.C) {
+	testRequires(c, DaemonIsLinux)
+	c.Assert(s.d.StartWithBusybox(), check.IsNil)
+
+	// the application is chosen so it generates output and doesn't react to SIGTERM
+	out, err := s.d.Cmd("run", "-d", "busybox", "sh", "-c", "while true;do date;done")
+	c.Assert(err, check.IsNil, check.Commentf("Output: %s", out))
+	id := strings.TrimSpace(out)
+	c.Assert(s.d.cmd.Process.Signal(os.Kill), check.IsNil)
+
+	// restart daemon.
+	if err := s.d.Restart(); err != nil {
+		c.Fatal(err)
+	}
+
+	out, err = s.d.Cmd("start", id)
+	c.Assert(err, check.IsNil, check.Commentf("Output: %s", out))
+}

libcontainerd/client_shutdownrestore_linux.go

@@ -20,6 +20,8 @@ func (clnt *client) Restore(containerID string, options ...CreateOption) error {
 		clnt.appendContainer(container)
 		clnt.unlock(cont.Id)
 
+		container.discardFifos()
+
 		if err := clnt.Signal(containerID, int(syscall.SIGTERM)); err != nil {
 			logrus.Errorf("error sending sigterm to %v: %v", containerID, err)
 		}
@@ -37,5 +39,8 @@ func (clnt *client) Restore(containerID string, options ...CreateOption) error {
 			return nil
 		}
 	}
+
+	clnt.deleteContainer(containerID)
+
 	return clnt.setExited(containerID)
 }

libcontainerd/container_linux.go

@@ -2,6 +2,7 @@ package libcontainerd
 
 import (
 	"encoding/json"
+	"io"
 	"io/ioutil"
 	"os"
 	"path/filepath"
@@ -191,3 +192,18 @@ func (ctr *container) handleEvent(e *containerd.Event) error {
 	}
 	return nil
 }
+
+// discardFifos attempts to fully read the container fifos to unblock processes
+// that may be blocked on the writer side.
+func (ctr *container) discardFifos() {
+	for _, i := range []int{syscall.Stdout, syscall.Stderr} {
+		f := ctr.fifo(i)
+		c := make(chan struct{})
+		go func() {
+			close(c) // this channel is used to not close the writer too early, before readonly open has been called.
+			io.Copy(ioutil.Discard, openReaderFromFifo(f))
+		}()
+		<-c
+		closeReaderFifo(f) // avoid blocking permanently on open if there is no writer side
+	}
+}