If you are looking to make Docker available on your favorite software distribution, this document is for you. It summarizes the requirements for building and running the Docker client and the Docker daemon.
We want to help you package Docker successfully. Before doing any packaging, a good first step is to introduce yourself on the docker-dev mailing list, explain what you're trying to achieve, and tell us how we can help. Don't worry, we don't bite! There might even be someone already working on packaging for the same distro!
You can also join the IRC channel - #docker and #docker-dev on Freenode are both active and friendly.
We like to refer to Tianon ("@tianon" on GitHub and "tianon" on IRC) as our "Packagers Relations", since he's always working to make sure our packagers have a good, healthy upstream to work with (both in our communication and in our build scripts). If you're having any kind of trouble, feel free to ping him directly. He also likes to keep track of what distributions we have packagers for, so feel free to reach out to him even just to say "Hi!"
If possible, your package should be called "docker". If that name is already taken, a second choice is "lxc-docker", but with the caveat that "LXC" is now an optional dependency (as noted below). Another possible choice is "docker.io".
The Docker project maintains its own build and release toolchain. It is pretty neat and entirely based on Docker (surprise!). This toolchain is the canonical way to build Docker. We encourage you to give it a try, and if the circumstances allow you to use it, we recommend that you do.
You might not be able to use the official build toolchain - usually because your distribution has a toolchain and packaging policy of its own. We get it! Your house, your rules. The rest of this document should give you the information you need to package Docker your way, without denaturing it in the process.
To build Docker, you will need the following:
- A recent version of git and mercurial
- Go version 1.3 or later
- A clean checkout of the source added to a valid Go
workspace under the path
src/github.com/docker/docker (unless you plan to use
AUTO_GOPATH
, explained in more detail below).
To build the Docker daemon, you will additionally need:
- An amd64/x86_64 machine running Linux
- SQLite version 3.7.9 or later
- libdevmapper version 1.02.68-cvs (2012-01-26) or later from lvm2 version 2.02.89 or later
- btrfs-progs version 3.8 or later (including commit e5cb128 from 2013-01-07) for the necessary btrfs headers
Be sure to also check out Docker's Dockerfile for the most up-to-date list of these build-time dependencies.
All Go dependencies are vendored under "./vendor". They are used by the official build, so the source of truth for the current version of each dependency is whatever is in "./vendor".
To use the vendored dependencies, simply make sure the path to "./vendor" is
included in GOPATH
(or use AUTO_GOPATH
, as explained below).
If you would rather (or must, due to distro policy) package these dependencies yourself, take a look at "./hack/vendor.sh" for an easy-to-parse list of the exact version for each.
NOTE: if you're not able to package the exact version (to the exact commit) of a given dependency, please get in touch so we can remediate! Who knows what discrepancies can be caused by even the slightest deviation. We promise to do our best to make everybody happy.
Please, please, please do not strip any compiled binaries. This is really important.
In our own testing, stripping the resulting binaries sometimes results in a binary that appears to work, but more often causes random panics, segfaults, and other issues. Even if the binary appears to work, please don't strip.
See the following quotes from Dave Cheney, which explain this position better from the upstream Golang perspective.
Super super important: Do not strip go binaries or archives. It isn't tested, often breaks, and doesn't work.
To quote myself: "Please do not strip Go binaries, it is not supported, not tested, is often broken, and doesn't do what you want"
To unpack that a bit
- not supported, as in, we don't support it, and recommend against it when asked
- not tested, we don't test stripped binaries as part of the build CI process
- is often broken, stripping a go binary will produce anywhere from no, to subtle, to outright execution failure, see above
To clarify my previous statements.
- I do not disagree with the debian policy, it is there for a good reason
- Having said that, it stripping Go binaries doesn't work, and nobody is looking at making it work, so there is that.
Thanks for patching the build formula.
Please use our build script ("./hack/make.sh") for all your compilation of Docker. If there's something you need that it isn't doing, or something it could be doing to make your life as a packager easier, please get in touch with Tianon and help us rectify the situation. Chances are good that other packagers have probably run into the same problems and a fix might already be in the works, but none of us will know for sure unless you harass Tianon about it. :)
All the commands listed within this section should be run with the Docker source checkout as the current working directory.
If you'd rather not be bothered with the hassles that setting up GOPATH
appropriately can be, and prefer to just get a "build that works", you should
add something similar to this to whatever script or process you're using to
build Docker:
export AUTO_GOPATH=1
This will cause the build scripts to set up a reasonable GOPATH
that
automatically and properly includes both docker/docker from the local
directory, and the local "./vendor" directory as necessary.
If you're building a binary that may need to be used on platforms that include
AppArmor, you will need to set DOCKER_BUILDTAGS
as follows:
export DOCKER_BUILDTAGS='apparmor'
If you're building a binary that may need to be used on platforms that include
SELinux, you will need to use the selinux
build tag:
export DOCKER_BUILDTAGS='selinux'
If your version of btrfs-progs is < 3.16.1 (also called btrfs-tools), then you will need the following tag to not check for btrfs version headers:
export DOCKER_BUILDTAGS='btrfs_noversion'
There are build tags for disabling graphdrivers as well. By default, support for all graphdrivers are built in.
To disable btrfs:
export DOCKER_BUILDTAGS='exclude_graphdriver_btrfs'
To disable devicemapper:
export DOCKER_BUILDTAGS='exclude_graphdriver_devicemapper'
To disable aufs:
export DOCKER_BUILDTAGS='exclude_graphdriver_aufs'
NOTE: if you need to set more than one build tag, space separate them:
export DOCKER_BUILDTAGS='apparmor selinux exclude_graphdriver_aufs'
If it is feasible within the constraints of your distribution, you should seriously consider packaging Docker as a single static binary. A good comparison is Busybox, which is often packaged statically as a feature to enable mass portability. Because of the unique way Docker operates, being similarly static is a "feature".
To build a static Docker daemon binary, run the following command (first ensuring that all the necessary libraries are available in static form for linking - see the "Build Dependencies" section above, and the relevant lines within Docker's own Dockerfile that set up our official build environment):
./hack/make.sh binary
This will create a static binary under "./bundles/$VERSION/binary/docker-$VERSION", where "$VERSION" is the contents of the file "./VERSION". This binary is usually installed somewhere like "/usr/bin/docker".
If you are only interested in a Docker client binary, set DOCKER_CLIENTONLY
to a non-empty value using something similar to the following: (which will prevent the extra step of compiling dockerinit)
export DOCKER_CLIENTONLY=1
If you need to (due to distro policy, distro library availability, or for other reasons) create a dynamically compiled daemon binary, or if you are only interested in creating a client binary for Docker, use something similar to the following:
./hack/make.sh dynbinary
This will create "./bundles/$VERSION/dynbinary/docker-$VERSION", which for client-only builds is the important file to grab and install as appropriate.
For daemon builds, you will also need to grab and install "./bundles/$VERSION/dynbinary/dockerinit-$VERSION", which is created from the minimal set of Docker's codebase that must be compiled statically (and is thus a pure static binary). The acceptable locations Docker will search for this file are as follows (in order):
- as "dockerinit" in the same directory as the daemon binary (ie, if docker is installed at "/usr/bin/docker", then "/usr/bin/dockerinit" will be the first place this file is searched for)
- "/usr/libexec/docker/dockerinit" or "/usr/local/libexec/docker/dockerinit" (FHS 3.0 Draft)
- "/usr/lib/docker/dockerinit" or "/usr/local/lib/docker/dockerinit" (FHS 2.3)
If (and please, only if) one of the paths above is insufficient due to distro
policy or similar issues, you may use the DOCKER_INITPATH
environment variable
at compile-time as follows to set a different path for Docker to search:
export DOCKER_INITPATH=/usr/lib/docker.io/dockerinit
If you find yourself needing this, please don't hesitate to reach out to Tianon to see if it would be reasonable or helpful to add more paths to Docker's list, especially if there's a relevant standard worth referencing (such as the FHS).
Also, it goes without saying, but for the purposes of the daemon please consider these two binaries ("docker" and "dockerinit") as if they were a single unit. Mixing and matching can cause undesired consequences, and will fail to run properly.
To function properly, the Docker daemon needs the following software to be installed and available at runtime:
- iptables version 1.4 or later
- procps (or similar provider of a "ps" executable)
- e2fsprogs version 1.4.12 or later (in use: mkfs.ext4, mkfs.xfs, tune2fs)
- XZ Utils version 4.9 or later
- a properly mounted cgroupfs hierarchy (having a single, all-encompassing "cgroup" mount point is not sufficient)
Additionally, the Docker client needs the following software to be installed and available at runtime:
- Git version 1.7 or later
The Docker daemon has very specific kernel requirements. Most pre-packaged kernels already include the necessary options enabled. If you are building your own kernel, you will either need to discover the options necessary via trial and error, or check out the Gentoo ebuild, in which a list is maintained (and if there are any issues or discrepancies in that list, please contact Tianon so they can be rectified).
Note that in client mode, there are no specific kernel requirements, and that the client will even run on alternative platforms such as Mac OS X / Darwin.
Some of Docker's features are activated by using optional command-line flags or by having support for them in the kernel or userspace. A few examples include:
- LXC execution driver (requires version 1.0 or later of the LXC utility scripts)
- AUFS graph driver (requires AUFS patches/support enabled in the kernel, and at least the "auplink" utility from aufs-tools)
- BTRFS graph driver (requires BTRFS support enabled in the kernel)
Docker expects to run as a daemon at machine startup. Your package will need to include a script for your distro's process supervisor of choice. Be sure to check out the "contrib/init" folder in case a suitable init script already exists (and if one does not, contact Tianon about whether it might be appropriate for your distro's init script to live there too!).
In general, Docker should be run as root, similar to the following:
docker -d
Generally, a DOCKER_OPTS
variable of some kind is available for adding more
flags (such as changing the graph driver to use BTRFS, switching the location of
"/var/lib/docker", etc).
As a final note, please do feel free to reach out to Tianon at any time for pretty much anything. He really does love hearing from our packagers and wants to make sure we're not being a "hostile upstream". As should be a given, we appreciate the work our packagers do to make sure we have broad distribution!