sysext-bakery: Recipes for baking systemd-sysext images

Systemd-sysext images are overlay images for /usr, allowing to extend the base OS with custom (static) binaries. Flatcar Container Linux as an OS without a package manager is a good fit for extension through systemd-sysext. The tools in this repository help you to create your own sysext images bundeling software to extend your base OS. The current focus is on Docker and containerd, contributions are welcome for other software.

Systemd-sysext

The NAME.raw sysext images (or NAME sysext directories) can be placed under /etc/extensions/ or /var/lib/extensions to be activated on boot by systemd-sysext.service. While systemd-sysext images are not really meant to also include the systemd service, Flatcar ships ensure-sysext.service as workaround to automatically load the image's services. This helper service is bound to systemd-sysext.service which activates the sysext images on boot. Currently it reloads the unit files from disk and reevaluates multi-user.target, sockets.target, and timers.target, making sure your enabled systemd units run. In the future systemd-sysext will only reload the unit files when this is upstream behavior (the current upstream behavior is to do nothing and leave it to the user). That means you need to use Upholds= drop-ins for the target units to start your units. At runtime executing systemctl restart systemd-sysext ensure-sysext will reload the sysext images and start the services. A manual systemd-sysext refresh is not recommended.

The compatibility mechanism of sysext images requires a metadata file in the image under usr/lib/extension-release.d/extension-release.NAME. It needs to contain a matching OS ID, and either a matching VERSION_ID or SYSEXT_LEVEL. Since the rapid release cycle and automatic updates of Flatcar Container Linux make it hard to rely on particular OS libraries by specifying a dependency of the sysext image to the OS version, it is not recommended to match by VERSION_ID. Instead, Flatcar defined the SYSEXT_LEVEL value 1.0 to match for. With systemd 252 you can also use ID=_any and then neither SYSEXT_LEVEL nor VERSION_ID are needed. The sysext image should only include static binaries.

Inside the image, binaries should be placed under usr/bin/ and systemd units under usr/lib/systemd/system/. While placing symlinks in the image itself to enable the units in the same way as systemd would normally do (like sockets.target.wants/my.socket → ../my.socket) is still currently supported, this is not a recommended practice. The recommended way is to ship drop-ins for the target units that start your unit. The drop-in file should use the Upholds= property in the [Unit] section. For example, for starting docker.socket we would use a drop-in for sockets.target placed in usr/lib/systemd/system/sockets.target.d/10-docker-socket.conf with the following contents:

[Unit]
Upholds=docker.socket

This can be done also for services, so for docker.service started by multi-user.target, the drop-in would reside in usr/lib/systemd/system/multi-user.target.d/10-docker-service.conf and it would have a Upholds=docker.service line instead.

The following Butane Config (YAML) can be be transpiled to Ignition JSON and will download a custom Docker+containerd sysext image on first boot. It also takes care of disabling Torcx and future inbuild Docker and containerd sysext images we plan to ship in Flatcar. If your sysext image doesn't replace Flatcar's inbuilt Docker/containerd, omit the two links entries and the torcx-generator entry.

variant: flatcar
version: 1.0.0
storage:
  files:
    - path: /etc/extensions/mydocker.raw
      contents:
        source: https://myserver.net/mydocker.raw
    - path: /etc/systemd/system-generators/torcx-generator
  links:
    - path: /etc/extensions/docker-flatcar.raw
      target: /dev/null
      overwrite: true
    - path: /etc/extensions/containerd-flatcar.raw
      target: /dev/null
      overwrite: true

Systemd-sysext on other distributions

The tools here will by default build for Flatcar and create the metadata file usr/lib/extension-release.d/extension-release.NAME as follows:

ID=flatcar
SYSEXT_LEVEL=1.0

This means other distributions will reject to load the sysext image by default. Use the configuration parameters in the tools to build for your distribution (pass OS= to be the OS ID from /etc/os-release) or to build for any distribution (pass OS=_any). You can also set the architecture to be arm64 to fetch the right binaries and encode this information in the sysext image metadata.

To add the automatic systemd unit loading to your distribution, store ensure-sysext.service in your systemd folder (e.g., /etc/systemd/system/) and enable the units: systemctl enable --now ensure-sysext.service systemd-sysext.service.

Recipes in this repository

The tools normally generate squashfs images not only because of the compression benefits but also because it doesn't need root permissions and loop device mounts.

Consuming the published images

There is a Github Action to build current recipes and to publish the built images as release artifacts. It's possible to directly consume the latest release from a Butane/Ignition configuration, example:

# butane < config.yaml > config.json
# ./flatcar_production_qemu.sh -i ./config.json
variant: flatcar
version: 1.0.0
storage:
  files:
    - path: /opt/extensions/docker/docker-24.0.5-x86-64.raw
      contents:
        source: https://github.com/flatcar/sysext-bakery/releases/download/20230901/docker-24.0.5-x86-64.raw
    - path: /opt/extensions/kubernetes/kubernetes-v1.27.4-x86-64.raw
      contents:
        source: https://github.com/flatcar/sysext-bakery/releases/download/20230901/kubernetes-v1.27.4-x86-64.raw
    - path: /etc/systemd/system-generators/torcx-generator
    - path: /etc/sysupdate.d/noop.conf
      contents:
        source: https://github.com/flatcar/sysext-bakery/releases/download/20230901/noop.conf
    - path: /etc/sysupdate.docker.d/docker.conf
      contents:
        source: https://github.com/flatcar/sysext-bakery/releases/download/20230901/docker.conf
    - path: /etc/sysupdate.kubernetes.d/kubernetes.conf
      contents:
        source: https://github.com/flatcar/sysext-bakery/releases/download/20230901/kubernetes.conf
  links:
    - target: /opt/extensions/docker/docker-24.0.5-x86-64.raw
      path: /etc/extensions/docker.raw
      hard: false
    - target: /opt/extensions/kubernetes/kubernetes-v1.27.4-x86-64.raw
      path: /etc/extensions/kubernetes.raw
      hard: false
    - path: /etc/extensions/docker-flatcar.raw
      target: /dev/null
      overwrite: true
    - path: /etc/extensions/containerd-flatcar.raw
      target: /dev/null
      overwrite: true
systemd:
  units:
    - name: systemd-sysupdate.timer
      enabled: true
    - name: systemd-sysupdate.service
      dropins:
        - name: docker.conf
          contents: |
            [Service]
            ExecStartPre=/usr/lib/systemd/systemd-sysupdate -C docker update
        - name: kubernetes.conf
          contents: |
            [Service]
            ExecStartPre=/usr/lib/systemd/systemd-sysupdate -C kubernetes update
        - name: sysext.conf
          contents: |
            [Service]
            ExecStartPost=systemctl restart systemd-sysext

This also configures systemd-sysupdate for auto-updates. The noop.conf is a workaround for systemd-sysupdate to run without error messages. Since the configuration sets up a custom Docker version, it also disables Torcx and the future docker-flatcar and containerd-flatcar extensions to prevent conflicts.

In the Flatcar docs you can find an Ignition configuration that explicitly sets the update configurations instead of downloading them.

The updates works by systemd-sysupdate fetching the SHA256SUMS file of the generated artifacts, which holds the list of built images with their respective SHA256 digest.

Creating a custom Docker sysext image

The Docker releases publish static binaries including containerd and the only missing piece are the systemd units. To ease the process, the create_docker_sysext.sh helper script takes care of downloading the release binaries and adding the systemd unit files, and creates a combined Docker+containerd sysext image:

./create_docker_sysext.sh 20.10.13 mydocker
[… writes mydocker.raw into current directory …]

Pass the OS or ARCH environment variables to build for another target than Flatcar amd64, e.g., for any distro with arm64:

OS=_any ARCH=arm64 ./create_docker_sysext.sh 20.10.13 mydocker
[… writes mydocker.raw into current directory …]

See the above intro section on how to use the resulting sysext image.

You can also limit the sysext image to only Docker (without containerd and runc) or only containerd (no Docker but runc) by passing the environment variables ONLY_DOCKER=1 or ONLY_CONTAINERD=1. If you build both sysext images that way, you can load both combined and, e.g., only load the Docker sysext image for debugging while using the containerd sysext image by default for Kubernetes.

Converting a Torcx image

Torcx was a solution for switching between different Docker versions on Flatcar. In case you have an existing Torcx image you can convert it with the convert_torcx_image.sh helper script (Currently only Torcx tar balls are supported and the conversion is done on best effort):

./convert_torcx_image.sh TORCXTAR SYSEXTNAME
[… writes SYSEXTNAME.raw into the current directory …]

Please make also sure that your don't have a containerd.service drop in file under /etc that uses Torcx paths.