moby/hack/dind

#!/bin/bash

# DinD: a wrapper script which allows docker to be run inside a docker container.
# Original version by Jerome Petazzoni <jerome@dotcloud.com>
# See the blog post: http://blog.docker.io/2013/09/docker-can-now-run-within-docker/
#
# This script should be executed inside a docker container in privilieged mode
# ('docker run -privileged', introduced in docker 0.6).

# Usage: dind CMD [ARG...]

# First, make sure that cgroups are mounted correctly.
CGROUP=/sys/fs/cgroup

[ -d $CGROUP ] || 
	mkdir $CGROUP

mountpoint -q $CGROUP || 
	mount -n -t tmpfs -o uid=0,gid=0,mode=0755 cgroup $CGROUP || {
		echo "Could not make a tmpfs mount. Did you use -privileged?"
		exit 1
	}

if [ -d /sys/kernel/security ] && ! mountpoint -q /sys/kernel/security
then
    mount -t securityfs none /sys/kernel/security || {
	echo "Could not mount /sys/kernel/security."
	echo "AppArmor detection and -privileged mode might break."
    }
fi

# Mount the cgroup hierarchies exactly as they are in the parent system.
for SUBSYS in $(cut -d: -f2 /proc/1/cgroup)
do
	[ -d $CGROUP/$SUBSYS ] || mkdir $CGROUP/$SUBSYS
	mountpoint -q $CGROUP/$SUBSYS || 
		mount -n -t cgroup -o $SUBSYS cgroup $CGROUP/$SUBSYS

	# The two following sections address a bug which manifests itself
	# by a cryptic "lxc-start: no ns_cgroup option specified" when
	# trying to start containers withina container.
	# The bug seems to appear when the cgroup hierarchies are not
	# mounted on the exact same directories in the host, and in the
	# container.

	# Named, control-less cgroups are mounted with "-o name=foo"
	# (and appear as such under /proc/<pid>/cgroup) but are usually
	# mounted on a directory named "foo" (without the "name=" prefix).
	# Systemd and OpenRC (and possibly others) both create such a
	# cgroup. To avoid the aforementioned bug, we symlink "foo" to
	# "name=foo". This shouldn't have any adverse effect.
	echo $SUBSYS | grep -q ^name= && {
		NAME=$(echo $SUBSYS | sed s/^name=//)
		ln -s $SUBSYS $CGROUP/$NAME
	}

	# Likewise, on at least one system, it has been reported that
	# systemd would mount the CPU and CPU accounting controllers
	# (respectively "cpu" and "cpuacct") with "-o cpuacct,cpu"
	# but on a directory called "cpu,cpuacct" (note the inversion
	# in the order of the groups). This tries to work around it.
	[ $SUBSYS = cpuacct,cpu ] && ln -s $SUBSYS $CGROUP/cpu,cpuacct
done

# Note: as I write those lines, the LXC userland tools cannot setup
# a "sub-container" properly if the "devices" cgroup is not in its
# own hierarchy. Let's detect this and issue a warning.
grep -q :devices: /proc/1/cgroup ||
	echo "WARNING: the 'devices' cgroup should be in its own hierarchy."
grep -qw devices /proc/1/cgroup ||
	echo "WARNING: it looks like the 'devices' cgroup is not mounted."

# Now, close extraneous file descriptors.
pushd /proc/self/fd >/dev/null
for FD in *
do
	case "$FD" in
	# Keep stdin/stdout/stderr
	[012])
		;;
	# Nuke everything else
	*)
		eval exec "$FD>&-"
		;;
	esac
done
popd >/dev/null

# Mount /tmp
mount -t tmpfs none /tmp

[ "$1" ] && exec "$@"
echo "You probably want to run hack/make.sh, or maybe a shell?"