PID tracking in modern init systems

Wherever your init daemon preference may go to, if you’re involved in writing daemons you’re currently faced with the following options:

Why is tracking daemon process IDs hard ? Because as a parent, you don’t have many options to watch how your children processes evolve, the only reliable way is to find a way for the children to send back their PID in some way.

Traditionally this has been done with PID files which are usually subject to being left. upstart has a mechanism to do tracking which is notorious for its ability to lose track of the real child PID and which leaves init in a complete screwed up state, with the original side-effect of hanging machines during shutdown.

Additionaly, to provide efficient sequencing of daemon startups, traditional init scripts resort to sleeping for arbitrary periods to ensure daemons are started while having no guarantee as to the actual readyness of the service.

In this article we’ll explore two ways of enabling daemons to coexist with upstart and systemd.

Upstart and expect stop

upstart has four modes of launching daemons:

When using expect stop, by tracking childs for SIGSTOP signals, upstart is able to reliably determine which PID the daemon lives under. When launched from upstart, the UPSTART_JOB environment variable is set, which means that it suffices to check for it. This also gives a good indication to the daemon that it should stay in the foreground:

const char    *upstart_job = getenv("UPSTART_JOB");

  if (upstart_job != NULL)
    raise(SIGSTOP); /* wait for upstart to start us up again */
  else {

    switch ((pid = fork())) {
    case -1:
      /* handle error */
    case 0:
      /* we're in the parent */
      return 0;


    if (open("/dev/null", O_RDWR) != 0)
      err(1, "cannot open /dev/null as stdin");
    if (dup(0) != 1)
      err(1, "cannot open /dev/null as stdout");
    if (dup(0) != 2)
      err(1, "cannot open /dev/null as stderr");

This is actually all there is to it as far as upstart is concerned. Upstart will catch the signal and register the PID it came from as the daemon. This way there is no risk of upstart losing track of the correct PID.

Let’s test our daemon manually:

$ env UPSTART_JOB=t $HOME/mydaemon
$ ps auxw | grep mydaemon
pyr      22702  0.0  0.0  22044  1576 ?        Ts   21:21   0:00 /home/pyr/mydaemon

The interesting bit here is that the reported state contains T for stopped. We can now resume execution by issuing:

kill -CONT 22702

Now configuring your daemon in upstart just needs:

expect stop
exec /home/pyr/mydaemon

Systemd and sd_notify

Systemd provides a similar facility for daemons, although it goes a bit further at the expense of an increased complexity.

Systemd’s approach is to expose a UNIX datagram socket to daemons for feedback purposes. The payload is composed of line separated key value pairs, where keys may be either one of:

A word of advice, if you plan on using the error notification mechanism, it would be advisable to pre-allocate a static buffer to be able to send out messages even in out-of-memory situations.

Like upstart, systemd sets an environment variable: NOTIFY_SOCKET to allow conditional behavior. It’s up to the daemon to include its PID in the message payload, so it doesn’t matter whether forking happens or not.

Systemd is happy with both forking and foreground-running daemons, a simple recommendation could be to only daemonize when neither UPSTART_JOB nor NOTIFY_SOCKET are present:

  const char    *upstart_job = getenv("UPSTART_JOB");
  const char    *systemd_socket = getenv("NOTIFY_SOCKET");

/* ... */

  if (upstart_job != NULL)
    raise(SIGSTOP); /* wait for upstart to start us up again */
  else if (notify_socket != NULL)
    sd_notify(0, "READY=1\nSTATUS=ready\nMAINPID=%lu\n",
    /* daemonize ... */

The use of sd_notify requires linking to libsystemd, if you want to keep dependencies to a minimum, you also have the possibility of crafting the payload directly and sending a single UDP datagram to the socket stored in the NOTIFY_SOCKET environment variable. Here’s an implementation from Vincent Bernat’s LLDPD:

To configure your systemd unit, you’ll now need to mark your job as having the type notify:

Description=My daemon