fix systemd-notify when using a different PID namespace

The current support of systemd-notify has a race condition as the
message send to the systemd notify socket might be dropped if the sender
process is not running by the time systemd checks for the sender of the
datagram.  A proper fix of this in systemd would require changes to the
kernel to maintain the cgroup of the sender process when it is dead (but
it is not probably going to happen...)
Generally, the solution to this issue is to specify the PID in the
message itself so that systemd has not to guess the sender, but this
wouldn't work when running in a PID namespace as the container will pass
the PID known in its namespace (something like PID=1,2,3..) and systemd
running on the host is not able to map it to the runc service.

The proposed solution is to have a proxy in runc that forwards the
messages to the host systemd.

Example of this issue:

https://github.com/projectatomic/atomic-system-containers/pull/24

Signed-off-by: Giuseppe Scrivano <gscrivan@redhat.com>

signals.go

@@ -17,7 +17,9 @@ const signalBufferSize = 2048
 
 // newSignalHandler returns a signal handler for processing SIGCHLD and SIGWINCH signals
 // while still forwarding all other signals to the process.
-func newSignalHandler(enableSubreaper bool) *signalHandler {
+// If notifySocket is present, use it to read systemd notifications from the container and
+// forward them to notifySocketHost.
+func newSignalHandler(enableSubreaper bool, notifySocket *notifySocket) *signalHandler {
 	if enableSubreaper {
 		// set us as the subreaper before registering the signal handler for the container
 		if err := system.SetSubreaper(1); err != nil {
@@ -30,7 +32,8 @@ func newSignalHandler(enableSubreaper bool) *signalHandler {
 	// handle all signals for the process.
 	signal.Notify(s)
 	return &signalHandler{
-		signals: s,
+		signals:      s,
+		notifySocket: notifySocket,
 	}
 }
 
@@ -42,7 +45,8 @@ type exit struct {
 }
 
 type signalHandler struct {
-	signals chan os.Signal
+	signals      chan os.Signal
+	notifySocket *notifySocket
 }
 
 // forward handles the main signal event loop forwarding, resizing, or reaping depending
@@ -54,6 +58,11 @@ func (h *signalHandler) forward(process *libcontainer.Process, tty *tty) (int, e
 	if err != nil {
 		return -1, err
 	}
+
+	if h.notifySocket != nil {
+		go h.notifySocket.run()
+	}
+
 	// perform the initial tty resize.
 	tty.resize()
 	for s := range h.signals {
@@ -75,6 +84,9 @@ func (h *signalHandler) forward(process *libcontainer.Process, tty *tty) (int, e
 					// status because we must ensure that any of the go specific process
 					// fun such as flushing pipes are complete before we return.
 					process.Wait()
+					if h.notifySocket != nil {
+						h.notifySocket.Close()
+					}
 					return e.status, nil
 				}
 			}