This recipe lets you run Docker within Docker.

There is only one requirement: your Docker version should support the --privileged flag.

If you came here because you would like to run a testing system like Jenkins in a container, and want that container to spin up more containers, then please read this blog post first. Thank you!

This work is now obsolete, thanks to the combined efforts of some amazing people like @jfrazelle and @tianon, who also are black belts in the art of putting IKEA furniture together.

If you want to run Docker-in-Docker today, all you need to do is:

docker run --privileged -d docker:dind

... And that's it; you get Docker running in Docker, thanks to the official Docker image, in its "Docker-in-Docker" flavor. You can then connect to this Docker instance by starting another Docker container linking to the first one (which is a pretty amazing thing to do).

For more details about the docker:dind official image, explanations about how to use it, customize it to use specific storage drivers, and other tidbits of useful knowledge, check its documentation on the Docker Hub.

... Then you're probably an archaelogist, a masochist, or both.

Seriously, though: the information below is here mostly for historical value, or if you want to understand how those things work under the hood.

You've been warned!

Build the image:

docker build -t dind .

Run Docker-in-Docker and get a shell where you can play, and docker daemon logs to stdout:

docker run --privileged -t -i dind

Run Docker-in-Docker and get a shell where you can play, but docker daemon logs into /var/log/docker.log:

docker run --privileged -t -i -e LOG=file dind

Run Docker-in-Docker and expose the inside Docker to the outside world:

docker run --privileged -d -p 4444 -e PORT=4444 dind

Note: when started with the PORT environment variable, the image will just the Docker daemon and expose it over said port. When started without the PORT environment variable, the image will run the Docker daemon in the background and execute a shell for you to play.

You can use the DOCKER_DAEMON_ARGS environment variable to configure the docker daemon with any extra options:

docker run --privileged -d -e DOCKER_DAEMON_ARGS="-D" dind

If you get a weird permission message, check the output of dmesg: it could be caused by AppArmor. In that case, try again, adding an extra flag to kick AppArmor out of the equation:

docker run --privileged --lxc-conf="lxc.aa_profile=unconfined" -t -i dind

If you get the warning:

WARNING: the 'devices' cgroup should be in its own hierarchy.

When starting up dind, you can get around this by shutting down docker and running:

# /etc/init.d/lxc stop
# umount /sys/fs/cgroup/
# mount -t cgroup devices 1 /sys/fs/cgroup

If the unmount fails, you can find out the proper mount-point with:

$ cat /proc/mounts | grep cgroup

The main trick is to have the --privileged flag. Then, there are a few things to care about:

• cgroups pseudo-filesystems have to be mounted, and they have to be mounted with the same hierarchies than the parent environment; this is done by a wrapper script, which is setup to run by default;
• /var/lib/docker cannot be on AUFS, so we make it a volume.

That's it.

Since AUFS cannot use an AUFS mount as a branch, it means that we have to use a volume. Therefore, all inner Docker data (images, containers, etc.) will be in the volume. Remember: volumes are not cleaned up when you docker rm, so if you wonder where did your disk space go after nesting 10 Dockers within each other, look no further :-)

Outside: it will use your installed version.

Inside: the Dockerfile will retrieve the latest docker binary from https://get.docker.io/; so if you want to include your own docker build, you will have to edit it. If you want to always use your local version, you could change the ADD line to be e.g.:

ADD /usr/bin/docker /usr/local/bin/docker

Yes. Note, however, that there seems to be a weird FD leakage issue. To work around it, the wrapdocker script carefully closes all the file descriptors inherited from the parent Docker and lxc-start (except stdio). I'm mentioning this in case you were relying on those inherited file descriptors, or if you're trying to repeat the experiment at home.

kojiromike/inception is a wrapper script that uses dind to nest Docker to arbitrary depth.

Also, when you will be exiting a nested Docker, this will happen:

root@975423921ac5:/# exit
root@6b2ae8bf2f10:/# exit
root@419a67dfdf27:/# exit
root@bc9f450caf22:/# exit
jpetazzo@tarrasque:~/Work/DOTCLOUD/dind$

At that point, you should blast Hans Zimmer's Dream Is Collapsing on your loudspeakers while twirling a spinning top.