Merge pull request #11246 from vrothberg/sdnotify=container

Implement SD-NOTIFY proxy in conmon

5 files changed, 76 insertions, 20 deletions

diff --git a/libpod/container.go b/libpod/container.go
index 80fd35c09..c57250d72 100644
--- a/libpod/container.go
+++ b/libpod/container.go
@@ -126,6 +126,10 @@ type Container struct {
 	// This is true if a container is restored from a checkpoint.
 	restoreFromCheckpoint bool
 
+	// Used to query the NOTIFY_SOCKET once along with setting up
+	// mounts etc.
+	notifySocket string
+
 	slirp4netnsSubnet *net.IPNet
 }
 
diff --git a/libpod/container_internal_linux.go b/libpod/container_internal_linux.go
index f21aebb09..8b73c82de 100644
--- a/libpod/container_internal_linux.go
+++ b/libpod/container_internal_linux.go
@@ -352,6 +352,10 @@ func (c *Container) generateSpec(ctx context.Context) (*spec.Spec, error) {
 		return nil, err
 	}
 
+	if err := c.mountNotifySocket(g); err != nil {
+		return nil, err
+	}
+
 	// Get host UID and GID based on the container process UID and GID.
 	hostUID, hostGID, err := butil.GetHostIDs(util.IDtoolsToRuntimeSpec(c.config.IDMappings.UIDMap), util.IDtoolsToRuntimeSpec(c.config.IDMappings.GIDMap), uint32(execUser.Uid), uint32(execUser.Gid))
 	if err != nil {
@@ -777,6 +781,41 @@ func (c *Container) generateSpec(ctx context.Context) (*spec.Spec, error) {
 	return g.Config, nil
 }
 
+// mountNotifySocket mounts the NOTIFY_SOCKET into the container if it's set
+// and if the sdnotify mode is set to container.  It also sets c.notifySocket
+// to avoid redundantly looking up the env variable.
+func (c *Container) mountNotifySocket(g generate.Generator) error {
+	notify, ok := os.LookupEnv("NOTIFY_SOCKET")
+	if !ok {
+		return nil
+	}
+	c.notifySocket = notify
+
+	if c.config.SdNotifyMode != define.SdNotifyModeContainer {
+		return nil
+	}
+
+	notifyDir := filepath.Join(c.bundlePath(), "notify")
+	logrus.Debugf("checking notify %q dir", notifyDir)
+	if err := os.MkdirAll(notifyDir, 0755); err != nil {
+		if !os.IsExist(err) {
+			return errors.Wrapf(err, "unable to create notify %q dir", notifyDir)
+		}
+	}
+	if err := label.Relabel(notifyDir, c.MountLabel(), true); err != nil {
+		return errors.Wrapf(err, "relabel failed %q", notifyDir)
+	}
+	logrus.Debugf("add bindmount notify %q dir", notifyDir)
+	if _, ok := c.state.BindMounts["/run/notify"]; !ok {
+		c.state.BindMounts["/run/notify"] = notifyDir
+	}
+
+	// Set the container's notify socket to the proxy socket created by conmon
+	g.AddProcessEnv("NOTIFY_SOCKET", "/run/notify/notify.sock")
+
+	return nil
+}
+
 // systemd expects to have /run, /run/lock and /tmp on tmpfs
 // It also expects to be able to write to /sys/fs/cgroup/systemd and /var/log/journal
 func (c *Container) setupSystemd(mounts []spec.Mount, g generate.Generator) error {
@@ -1730,6 +1769,7 @@ rootless=%d
 			c.state.BindMounts[dest] = src
 		}
 	}
+
 	return nil
 }
 
diff --git a/libpod/diff.go b/libpod/diff.go
index cdd5e79cb..6a50bef32 100644
--- a/libpod/diff.go
+++ b/libpod/diff.go
@@ -14,6 +14,7 @@ var initInodes = map[string]bool{
 	"/etc/resolv.conf":   true,
 	"/proc":              true,
 	"/run":               true,
+	"/run/notify":        true,
 	"/run/.containerenv": true,
 	"/run/secrets":       true,
 	"/sys":               true,
diff --git a/libpod/oci_conmon_exec_linux.go b/libpod/oci_conmon_exec_linux.go
index 05a4e19b0..469bc7d86 100644
--- a/libpod/oci_conmon_exec_linux.go
+++ b/libpod/oci_conmon_exec_linux.go
@@ -462,7 +462,7 @@ func (r *ConmonOCIRuntime) startExec(c *Container, sessionID string, options *Ex
 		Setpgid: true,
 	}
 
-	err = startCommandGivenSelinux(execCmd)
+	err = startCommandGivenSelinux(execCmd, c)
 
 	// We don't need children pipes  on the parent side
 	errorhandling.CloseQuiet(childSyncPipe)
diff --git a/libpod/oci_conmon_linux.go b/libpod/oci_conmon_linux.go
index 846d3815a..ff25be234 100644
--- a/libpod/oci_conmon_linux.go
+++ b/libpod/oci_conmon_linux.go
@@ -364,11 +364,6 @@ func (r *ConmonOCIRuntime) StartContainer(ctr *Container) error {
 		return err
 	}
 	env := []string{fmt.Sprintf("XDG_RUNTIME_DIR=%s", runtimeDir)}
-	if ctr.config.SdNotifyMode == define.SdNotifyModeContainer {
-		if notify, ok := os.LookupEnv("NOTIFY_SOCKET"); ok {
-			env = append(env, fmt.Sprintf("NOTIFY_SOCKET=%s", notify))
-		}
-	}
 	if path, ok := os.LookupEnv("PATH"); ok {
 		env = append(env, fmt.Sprintf("PATH=%s", path))
 	}
@@ -1014,12 +1009,6 @@ func (r *ConmonOCIRuntime) createOCIContainer(ctr *Container, restoreOptions *Co
 		}
 	}
 
-	if ctr.config.SdNotifyMode == define.SdNotifyModeIgnore {
-		if err := os.Unsetenv("NOTIFY_SOCKET"); err != nil {
-			logrus.Warnf("Error unsetting NOTIFY_SOCKET %v", err)
-		}
-	}
-
 	pidfile := ctr.config.PidFile
 	if pidfile == "" {
 		pidfile = filepath.Join(ctr.state.RunDir, "pidfile")
@@ -1027,6 +1016,10 @@ func (r *ConmonOCIRuntime) createOCIContainer(ctr *Container, restoreOptions *Co
 
 	args := r.sharedConmonArgs(ctr, ctr.ID(), ctr.bundlePath(), pidfile, ctr.LogPath(), r.exitsDir, ociLog, ctr.LogDriver(), logTag)
 
+	if ctr.config.SdNotifyMode == define.SdNotifyModeContainer && ctr.notifySocket != "" {
+		args = append(args, fmt.Sprintf("--sdnotify-socket=%s", ctr.notifySocket))
+	}
+
 	if ctr.config.Spec.Process.Terminal {
 		args = append(args, "-t")
 	} else if ctr.config.Stdin {
@@ -1171,7 +1164,8 @@ func (r *ConmonOCIRuntime) createOCIContainer(ctr *Container, restoreOptions *Co
 		}
 	}
 
-	err = startCommandGivenSelinux(cmd)
+	err = startCommandGivenSelinux(cmd, ctr)
+
 	// regardless of whether we errored or not, we no longer need the children pipes
 	childSyncPipe.Close()
 	childStartPipe.Close()
@@ -1203,7 +1197,13 @@ func (r *ConmonOCIRuntime) createOCIContainer(ctr *Container, restoreOptions *Co
 		// conmon not having a pid file is a valid state, so don't set it if we don't have it
 		logrus.Infof("Got Conmon PID as %d", conmonPID)
 		ctr.state.ConmonPID = conmonPID
-		if ctr.config.SdNotifyMode != define.SdNotifyModeIgnore {
+
+		// Send the MAINPID via sdnotify if needed.
+		switch ctr.config.SdNotifyMode {
+		case define.SdNotifyModeContainer, define.SdNotifyModeIgnore:
+		// Nothing to do or conmon takes care of it already.
+
+		default:
 			if sent, err := daemon.SdNotify(false, fmt.Sprintf("MAINPID=%d", conmonPID)); err != nil {
 				logrus.Errorf("Error notifying systemd of Conmon PID: %v", err)
 			} else if sent {
@@ -1239,11 +1239,6 @@ func (r *ConmonOCIRuntime) configureConmonEnv(ctr *Container, runtimeDir string)
 	}
 
 	extraFiles := make([]*os.File, 0)
-	if ctr.config.SdNotifyMode == define.SdNotifyModeContainer {
-		if notify, ok := os.LookupEnv("NOTIFY_SOCKET"); ok {
-			env = append(env, fmt.Sprintf("NOTIFY_SOCKET=%s", notify))
-		}
-	}
 	if !r.sdNotify {
 		if listenfds, ok := os.LookupEnv("LISTEN_FDS"); ok {
 			env = append(env, fmt.Sprintf("LISTEN_FDS=%s", listenfds), "LISTEN_PID=1")
@@ -1335,7 +1330,23 @@ func (r *ConmonOCIRuntime) sharedConmonArgs(ctr *Container, cuuid, bundlePath, p
 
 // startCommandGivenSelinux starts a container ensuring to set the labels of
 // the process to make sure SELinux doesn't block conmon communication, if SELinux is enabled
-func startCommandGivenSelinux(cmd *exec.Cmd) error {
+func startCommandGivenSelinux(cmd *exec.Cmd, ctr *Container) error {
+	// Make sure to unset the NOTIFY_SOCKET and reset if afterwards if needed.
+	switch ctr.config.SdNotifyMode {
+	case define.SdNotifyModeContainer, define.SdNotifyModeIgnore:
+		if ctr.notifySocket != "" {
+			if err := os.Unsetenv("NOTIFY_SOCKET"); err != nil {
+				logrus.Warnf("Error unsetting NOTIFY_SOCKET %v", err)
+			}
+
+			defer func() {
+				if err := os.Setenv("NOTIFY_SOCKET", ctr.notifySocket); err != nil {
+					logrus.Errorf("Error resetting NOTIFY_SOCKET=%s", ctr.notifySocket)
+				}
+			}()
+		}
+	}
+
 	if !selinux.GetEnabled() {
 		return cmd.Start()
 	}