Docker demons: PID-1, orphans, zombies, and signals.

There are a number of corner cases to consider when dealing with Docker, multiple processes, and signals. Probably the most famous post on this matter is from the Phusion blog. Here, we'll see some examples of how to see these problems first hand, and one way to work around it: fpco/pid1.

The Phusion blog post recommends using their baseimage-docker. This image provides a my_init entrypoint which handles the problems described here, as well as introducing some extra OS features, such as syslog handling. Unfortunately, we ran into problems with Phusion's usage of syslog-ng, in particular with it creating unkillable processes pegged at 100% CPU usage. We're still investigating the root cause, but in practice we have found that the syslog usage is a far less motivating case than simply a good init process, which is why we've created the pid1 Haskell package together with a simple fpco/pid1 Docker image.

This blog post is intended to be interactive: you'll get the most bang for your buck by opening up your terminal and running commands along with reading the text. It will be far more motivating to see your Ctrl-C completely fail to kill a process.

NOTE The primary reason we wrote our own implementation in Haskell was to be able to embed it within the Stack build tool. There are other lightweight init processes already available, such as dumb-init. I've also blogged about using dumb-init. While this post uses pid1, there's nothing specific to it versus other init processes.

Playing with entrypoints

Docker has a concept of entrypoints, which provides a default wrapping command for commands you provides to docker run. For example, consider this interaction with Docker:

$ docker run --entrypoint /usr/bin/env ubuntu:16.04 FOO=BAR bash c 'echo $FOO'
BAR

This works because the above is equivalent to:

$ docker run ubuntu:16.04 /usr/bin/env FOO=BAR bash -c 'echo $FOO'

Entrypoints can be overridden on the command line (as we just did), but can also be specified in the Dockerfile (which we'll do later). The default entrypoint for the ubuntu Docker image is a null entrypoint, meaning that the provided command will be run directly without any wrapping. We're going to simulate that experience by using /usr/bin/env as an entrypoint, since switching entrypoint back to null isn't yet supported in released Docker. When you run /usr/bin/env foo bar baz, the env process will exec the foo command, making foo the new PID 1, which for our purposes gives it the same behavior as a null entrypoint.

Both the fpco/pid1 and snoyberg/docker-testing images we'll use below set /sbin/pid1 as the default entrypoint. In the example commands, we're explicitly including --entrypoint /sbin/pid1. This is just to be clear on which entrypoint is being used; if you exclude that option, the same behavior will persist.

Sending TERM signal to process

We'll start with our sigterm.hs program, which runs ps (we'll see why soon), then sends itself a SIGTERM and then loops forever. On a Unix system, the default process behavior when receiving a SIGTERM is to exit. Therefore, we'd expect that our process will just exit when run. Let's see:

$ docker run --rm --entrypoint /usr/bin/env snoyberg/docker-testing sigterm
  PID TTY          TIME CMD
    1 ?        00:00:00 sigterm
    9 ?        00:00:00 ps
Still alive!
Still alive!
Still alive!
^C
$

The process ignored the SIGTERM and kept running, until I hit Ctrl-C (we'll see what that does later). Another feature in the sigterm code base, though, is that if you give it the command line argument install-handler, it will explicitly install a SIGTERM handler which will kill the process. Perhaps surprisingly, this has a significant impact on our application:

$ docker run --rm --entrypoint /usr/bin/env snoyberg/docker-testing sigterm install-handler
  PID TTY          TIME CMD
    1 ?        00:00:00 sigterm
    8 ?        00:00:00 ps
Still alive!
$

The reason for this is some Linux kernel magic: the kernel treats a process with PID 1 specially, and does not, by default, kill the process when receiving the SIGTERM or SIGINT signals. This can be very surprising behavior. For a simpler example, try running the following commands in two different terminals:

$ docker run --rm --name sleeper ubuntu:16.04 sleep 100
$ docker kill -s TERM sleeper

Notice how the docker run command does not exit, and if you check your ps aux output, you'll see that the process is still running. That's because the sleep process was not designed to be PID 1, and does not install a special signal handler. To work around this problem, you've got two choices:

1. Ensure every command you run from docker run has explicit handling of SIGTERM.
2. Make sure the command you run isn't PID 1, but instead use a process that is designed to handle SIGTERM correctly.

Let's see how the sigterm program works with our /sbin/pid1 entrypoint:

$ docker run --rm --entrypoint /sbin/pid1 snoyberg/docker-testing sigterm
  PID TTY          TIME CMD
    1 ?        00:00:00 pid1
    8 ?        00:00:00 sigterm
   12 ?        00:00:00 ps

The program exits immediately, as we'd like. But look at the ps output: our first process is now pid1 instead of sigterm. Since sigerm is being launched as a different PID (8 in this case), the special casing from the Linux kernel does not come into play, and default SIGTERM handling is active. To step through exactly what happens in our case:

1. Our container is created, and the command /usr/sbin/pid1 sigterm is run inside of it.
2. pid1 starts as PID-1, does its business, and then fork/execs the sigterm executable.
3. sigterm raises the SIGTERM signal to itself, causing it to die.
4. pid1 sees that its child died from SIGTERM (== signal 15) and exits with exit code 143 (== 128 + 15).
5. Since our PID1 is dead, our container dies too.

This isn't just some magic with sigterm, you can do the same thing with sleep:

$ docker run --rm --name sleeper fpco/pid1 sleep 100
$ docker kill -s TERM sleeper

Unlike with the ubuntu image, this will kill the container immediately, due to the /sbin/pid1 entrypoint used by fpco/pid1.

NOTE In the case of sigterm, which sends the TERM signal to itself, it turns out you don't need a special PID1 process with signal handling, anything will do. For example, try docker run --rm --entrypoint /usr/bin/env snoyberg/docker-testing /bin/bash -c "sigterm;echo bye". But playing with sleep will demonstrate the need for a real signal-aware PID1 process.

Ctrl-C: sigterm vs sleep

There's a slight difference between sigterm and sleep when it comes to the behavior of sending hitting Ctrl-C. When you use Ctrl-C, it sends a SIGINT to the docker run process, which proxies that signal to the process inside the container. sleep will ignore it, just as it ignores SIGTERM, due to the default signal handlers for PID1 in the Linux kernel. However, the sigterm executable is written in Haskell, and the Haskell runtime itself installs a signal handler that converts SIGINT into a user interrupt exception, overriding the PID1 default behavior. For more on signal proxying, see the docker attach documentation.

Reaping orphans

Suppose you have process A, which fork/execs process B. When process B dies, process A must call waitpid to get its exit status from the kernel, and until it does so, process B will be dead but with an entry in the system process table. This is known as being a zombie.

But what happens if process B outlives process A? In this case, process B is known as an orphan, and needs to be adopted by the init process, aka PID1. It is the init process's job to reap orphans so they do not remain as zombies.

The orphans.hs program will:

• Spawn a child process, and then loop forever calling ps
• In the child process: run the echo command a few times, without calling waitpid, and then exit

As you can see, none of the processes involved will reap the zombie echo processes. The output from the process confirms that we have, in fact, created zombies:

$ docker run --rm --entrypoint /usr/bin/env snoyberg/docker-testing orphans
1
2
3
4
Still alive!
  PID TTY          TIME CMD
    1 ?        00:00:00 orphans
    8 ?        00:00:00 orphans
   13 ?        00:00:00 echo <defunct>
   14 ?        00:00:00 echo <defunct>
   15 ?        00:00:00 echo <defunct>
   16 ?        00:00:00 echo <defunct>
   17 ?        00:00:00 ps
Still alive!
  PID TTY          TIME CMD
    1 ?        00:00:00 orphans
   13 ?        00:00:00 echo <defunct>
   14 ?        00:00:00 echo <defunct>
   15 ?        00:00:00 echo <defunct>
   16 ?        00:00:00 echo <defunct>
   18 ?        00:00:00 ps
Still alive!

And so on until we kill the container. That <defunct> indicates a zombie process. The issue is that our PID 1, orphans, doesn't do reaping. As you probably guessed, we can solve this by just using the /sbin/pid1 entrypoint:

$ docker run --rm --entrypoint /sbin/pid1 snoyberg/docker-testing orphans
1
2
3
4
Still alive!
  PID TTY          TIME CMD
    1 ?        00:00:00 pid1
   10 ?        00:00:00 orphans
   14 ?        00:00:00 orphans
   19 ?        00:00:00 echo <defunct>
   20 ?        00:00:00 echo <defunct>
   21 ?        00:00:00 echo <defunct>
   22 ?        00:00:00 echo <defunct>
   23 ?        00:00:00 ps
Still alive!
  PID TTY          TIME CMD
    1 ?        00:00:00 pid1
   10 ?        00:00:00 orphans
   24 ?        00:00:00 ps
Still alive!

pid1 now adopts the echo processes when the child orphans process dies, and reaps accordingly.

Surviving children

Let's try out something else: process A is the primary command for the Docker container, and it spawns process B. Before process B exits, process A exits, causing the Docker container to exit. In this case, the running process B will be forcibly closed by the kernel (see this Stack Overflow question for details). We can see this with our surviving.hs program

$ docker run --rm --entrypoint /usr/bin/env snoyberg/docker-testing surviving
Parent sleeping
Child: 1
Child: 2
Child: 4
Child: 3
Child: 1
Child: 2
Child: 3
Child: 4
Parent exiting

Unfortunately this doesn't give our child processes a chance to do any cleanup. Instead, we would rather send them a SIGTERM, and after a grace period send them a SIGKILL. This is exactly what pid1 does:

$ docker run --rm --entrypoint /sbin/pid1 snoyberg/docker-testing surviving
Parent sleeping
Child: 2
Child: 3
Child: 1
Child: 4
Child: 2
Child: 1
Child: 4
Child: 3
Parent exiting
Got a TERM
Got a TERM
Got a TERM
Got a TERM

Signaling docker run vs PID1

When you run sleep 60 and then hit Ctrl-C, the sleep process itself receives a SIGINT. When you instead run docker run --rm fpco/pid1 sleep 60 and hit Ctrl-C, you may think that the same thing is happening. However, in reality, it's not at all the same. Your docker run call creates a docker run process, which sends a command to the Docker daemon on your machine, and that daemon creates the actual sleep process (inside a container). When you hit Ctrl-C on your terminal, you're sending SIGINT to docker run, which is in fact sending a command to the Docker daemon, which in turn sends a SIGINT to your sleep process.

Want proof? Try out the following:

$ docker run --rm fpco/pid1 sleep 60&
[1] 417
$ kill -KILL $!
$ docker ps
CONTAINER ID        IMAGE                       COMMAND                  CREATED             STATUS              PORTS               NAMES
69fbc70e95e2        fpco/pid1                   "/sbin/pid1 sleep 60"    11 seconds ago      Up 11 seconds                           hopeful_mayer
[1]+  Killed                  docker run --rm fpco/pid1 sleep 60

In this case, we sent a SIGKILL to the docker run command. Unlike SIGINT or SIGTERM, and SIGKILL cannot be handled, and therefore docker run is unable to delegate signal handling to a different process. As a result, the docker run command itself dies, but the sleep process (and its container) continue running.

Some takeaways from this:

• Make sure you use something like pid1 so that your SIGINT or SIGTERM to the docker run process actually get your container to reliably shut down
• If you must send a SIGKILL to your process, use the docker kill command instead

Alternative to entrypoint

We've used --entrypoint /sbin/pid1 a lot here. In fact, each usage of that has been superfluous, since the fpco/pid1 and snoyberg/docker-testing images both use /sbin/pid1 as their default entrypoint anyway. I included it for explicitness. To prove it to you:

$ docker run --rm fpco/pid1 sleep 60
^C$

But if you don't want to muck with entrypoints, you can always just include /sbin/pid1 at the beginning of your command, e.g.:

$ docker run --rm --entrypoint /usr/bin/env fpco/pid1 /sbin/pid1 sleep 60
^C$

And if you have your own Docker image and you'd just like to include the pid1 executable, you can download it from the Github releases page.

Dockerfiles, command vs exec form

You may be tempted to put something like ENTRYPOINT /sbin/pid1 in your Dockerfile. Let's see why that won't work:

$ cat Dockerfile
FROM fpco/pid1
ENTRYPOINT /sbin/pid1
$ docker build --tag test .
Sending build context to Docker daemon 2.048 kB
Step 1 : FROM fpco/pid1
 ---> aef1f7b702b9
Step 2 : ENTRYPOINT /sbin/pid1
 ---> Using cache
 ---> f875b43a9e40
Successfully built f875b43a9e40
$ docker run --rm test ps
pid1: No arguments provided

The issue here is that we specified /sbin/pid1 in what Docker calls command form. This is just a raw string which is interpreted by the shell. It is unable to be passed an additional command (like ps), and therefore pid1 itself complains that it hasn't been told what to run. The correct way to specify your entrypoint is ENTRYPOINT ["/sbin/pid1"], e.g.:

$ cat Dockerfile
FROM fpco/pid1
ENTRYPOINT ["/sbin/pid1"]
$ docker build --tag test .
Sending build context to Docker daemon 2.048 kB
Step 1 : FROM fpco/pid1
 ---> aef1f7b702b9
Step 2 : ENTRYPOINT /sbin/pid1
 ---> Running in ba0fa8c5bd41
 ---> 4835dec4aae6
Removing intermediate container ba0fa8c5bd41
Successfully built 4835dec4aae6
$ docker run --rm test ps
  PID TTY          TIME CMD
    1 ?        00:00:00 pid1
    8 ?        00:00:00 ps

Generally speaking, you should stick with command form in your Dockerfiles at all times. It is explicit about whitespace handling, and avoids the need to use a shell as an interpreter.

Takeaways

The main takeaway here is: unless you have a good reason to do otherwise, you should use a minimal init process like pid1. The Phusion/my_init approach works, but may be too heavy weight for some. If you don't need syslog and other add-on features of Phusion, you're probably best with a minimal init instead.

As a separate but somewhat related comment: we're going to have a follow up post on this blog in the coming days explaining how we compiled the pid1 executable as a static executable to make it compatible with all various Linux flavors, and how you can do the same for your Haskell executables. Stay tuned!