Skip to content

Latest commit

 

History

History
295 lines (243 loc) · 14.5 KB

yaml.md

File metadata and controls

295 lines (243 loc) · 14.5 KB

Configuration Reference

The linuxkit build command assembles a set of containerised components into in image. The simplest type of image is just a tar file of the contents (useful for debugging) but more useful outputs add a Dockerfile to build a container, or build a full disk image that can be booted as a linuxKit VM. The main use case is to build an assembly that includes containerd to run a set of containers, but the tooling is very generic.

The yaml configuration specifies the components used to build up an image . All components are downloaded at build time to create an image. The image is self-contained and immutable, so it can be tested reliably for continuous delivery.

Components are specified as Docker images which are pulled from a registry during build if they are not available locally. See image-cache for more details on local caching. The Docker images are optionally verified with Docker Content Trust. For private registries or private repositories on a registry credentials provided via docker login are re-used.

The configuration file is processed in the order kernel, init, onboot, onshutdown, services, files. Each section adds files to the root file system. Sections may be omitted.

Each container that is specified is allocated a unique uid and gid that it may use if it wishes to run as an isolated user (or user namespace). Anywhere you specify a uid or gid field you specify either the numeric id, or if you use a name it will refer to the id allocated to the container with that name.

services:
  - name: redis
    image: redis:latest
    uid: redis
    gid: redis
    binds:
     - /etc/redis:/etc/redis
files:
  - path: /etc/redis/redis.conf
    contents: "..."
    uid: redis
    gid: redis
    mode: "0600"

kernel

The kernel section is only required if booting a VM. The files will be put into the boot/ directory, where they are used to build bootable images.

The kernel section defines the kernel configuration. The image field specifies the Docker image, which should contain a kernel file that will be booted (eg a bzImage for amd64) and a file called kernel.tar which is a tarball that is unpacked into the root, which should usually contain a kernel modules directory. cmdline specifies the kernel command line options if required.

To override the names, you can specify the kernel image name with binary: bzImage and the tar image with tar: kernel.tar or the empty string or none if you do not want to use a tarball at all.

Kernel packages may also contain a cpio archive containing CPU microcode which needs prepending to the initrd. To select this option, recommended when booting on bare metal, add ucode: intel-ucode.cpio to the kernel section.

init

The init section is a list of images that are used for the init system and are unpacked directly into the root filesystem. This should bring up containerd, start the system and daemon containers, and set up basic filesystem mounts. in the case of a LinuxKit system. For ease of modification runc and containerd images, which just contain these programs are added here rather than bundled into the init container.

onboot

The onboot section is a list of images. These images are run before any other images. They are run sequentially and each must exit before the next one is run. These images can be used to configure one shot settings. See Image specification for a list of supported fields.

onshutdown

This is a list of images to run on a clean shutdown. Note that you must not rely on these being run at all, as machines may be be powered off or shut down without having time to run these scripts. If you add anything here you should test both in the case where they are run and when they are not. Most systems are likely to be "crash only" and not have any setup here, but you can attempt to deregister cleanly from a network service here, rather than relying on timeouts, for example.

services

The services section is a list of images for long running services which are run with containerd. Startup order is undefined, so containers should wait on any resources, such as networking, that they need. See Image specification for a list of supported fields.

files

The files section can be used to add files inline in the config, or from an external file.

files:
  - path: dir
    directory: true
    mode: "0777"
  - path: dir/name1
    source: "/some/path/on/local/filesystem"
    mode: "0666"
  - path: dir/name2
    source: "/some/path/that/it/is/ok/to/omit"
    optional: true
    mode: "0666"
  - path: dir/name3
    contents: "orange"
    mode: "0644"
    uid: 100
    gid: 100

Specifying the mode is optional, and will default to 0600. Leading directories will be created if not specified. You can use ~/path in source to specify a path in the build user's home directory.

In addition there is a metadata option that will generate the file. Currently the only value supported here is "yaml" which will output the yaml used to generate the image into the specified file:

  - path: etc/linuxkit.yml
    metadata: yaml

Because a tmpfs is mounted onto /var, /run, and /tmp by default, the tmpfs mounts will shadow anything specified in files section for those directories.

Image specification

Entries in the onboot and services sections specify an OCI image and options. Default values may be specified using the org.mobyproject.config image label. For more details see the OCI specification.

If the org.mobylinux.config label is set in the image, that specifies default values for these fields if they are not set in the yaml file. While most fields are replaced if they are specified in the yaml file, some support add via the format <field>.add; see below. You can override the label entirely by setting the value, or setting it to be empty to remove the specification for that value in the label.

If you need an OCI option that is not specified here please open an issue or pull request as the list is not yet complete.

By default the containers will be run in the host net, ipc and uts namespaces, as that is the usual requirement; in many ways they behave like pods in Kubernetes. Mount points must already exist, as must a file or directory being bind mounted into a container.

  • name a unique name for the program being executed, used as the containerd id.
  • image the Docker image to use for the root filesystem. The default command, path and environment are extracted from this so they need not be filled in.
  • capabilities the Linux capabilities required, for example CAP_SYS_ADMIN. If there is a single capability all then all capabilities are added.
  • capabilities.add the Linux capabilities required, but these are added to the defaults, rather than overriding them.
  • ambient the Linux ambient capabilities (capabilities passed to non root users) that are required.
  • mounts is the full form for specifying a mount, which requires type, source, destination and a list of options. If any fields are omitted, sensible defaults are used if possible, for example if the type is dev it is assumed you want to mount at /dev. The default mounts and their options can be replaced by specifying a mount with new options here at the same mount point.
  • binds is a simpler interface to specify bind mounts, accepting a string like /src:/dest:opt1,opt2 similar to the -v option for bind mounts in Docker.
  • binds.add is a simpler interface to specify bind mounts, but these are added to the defaults, rather than overriding them.
  • tmpfs is a simpler interface to mount a tmpfs, like --tmpfs in Docker, taking /dest:opt1,opt2.
  • command will override the command and entrypoint in the image with a new list of commands.
  • env will override the environment in the image with a new environment list. Specify variables as VAR=value.
  • cwd will set the working directory, defaults to /.
  • net sets the network namespace, either to a path, or if none or new is specified it will use a new namespace.
  • ipc sets the ipc namespace, either to a path, or if new is specified it will use a new namespace.
  • uts sets the uts namespace, either to a path, or if new is specified it will use a new namespace.
  • pid sets the pid namespace, either to a path, or if host is specified it will use the host namespace.
  • readonly sets the root filesystem to read only, and changes the other default filesystems to read only.
  • maskedPaths sets paths which should be hidden.
  • readonlyPaths sets paths to read only.
  • uid sets the user id of the process.
  • gid sets the group id of the process.
  • additionalGids sets a list of additional groups for the process.
  • noNewPrivileges is true means no additional capabilities can be acquired and suid binaries do not work.
  • hostname sets the hostname inside the image.
  • oomScoreAdj changes the OOM score.
  • rootfsPropagation sets the rootfs propagation, eg shared, slave or (default) private.
  • cgroupsPath sets the path for cgroups.
  • resources sets cgroup resource limits as per the OCI spec.
  • sysctl sets a map of sysctl key value pairs that are set inside the container namespace.
  • rmlimits sets a list of rlimit values in the form name,soft,hard, eg nofile,100,200. You can use unlimited as a value too.
  • annotations sets a map of key value pairs as OCI metadata.

There are experimental userns, uidMappings and gidMappings options for user namespaces but these are not yet supported, and may have permissions issues in use.

In addition to the parts of the specification above used to generate the OCI spec, there is a runtime section in the image specification which specifies some actions to take place when the container is being started.

  • cgroups takes a list of cgroups that will be created before the container is run.
  • mounts takes a list of mount specifications (source, destination, type, options) and mounts them in the root namespace before the container is created. It will try to make any missing destination directories.
  • mkdir takes a list of directories to create at runtime, in the root mount namespace. These are created before the container is started, so they can be used to create directories for bind mounts, for example in /tmp or /run which would otherwise be empty.
  • interface defines a list of actions to perform on a network interface:
    • name specifies the name of an interface. An existing interface with this name will be moved into the container's network namespace.
    • add specifies a type of interface to be created in the containers namespace, with the specified name.
    • createInRoot is a boolean which specifes that the interface being added should be created in the root namespace first, then moved. This is needed for wireguard interfaces.
    • peer specifies the name of the other end when creating a veth interface. This end will remain in the root namespace, where it can be attached to a bridge. Specifying this implies add: veth.
  • bindNS specifies a namespace type and a path where the namespace from the container being created will be bound. This allows a namespace to be set up in an onboot container, and then using net: path for a service container to use that network namespace later.
  • namespace overrides the LinuxKit default containerd namespace to put the container in; only applicable to services.

An example of using the runtime config to configure a network namespace with wireguard and then run nginx in that namespace is shown below:

onboot:
  - name: dhcpcd
    image: linuxkit/dhcpcd:<hash>
    command: ["/sbin/dhcpcd", "--nobackground", "-f", "/dhcpcd.conf", "-1"]
  - name: wg
    image: linuxkit/ip:<hash>
    net: new
    binds:
      - /etc/wireguard:/etc/wireguard
    command: ["sh", "-c", "ip link set dev wg0 up; ip address add dev wg0 192.168.2.1 peer 192.168.2.2; wg setconf wg0 /etc/wireguard/wg0.conf; wg show wg0"]
    runtime:
      interfaces:
        - name: wg0
          add: wireguard
          createInRoot: true
      bindNS:
        net: /run/netns/wg
services:
  - name: nginx
    image: nginx:alpine
    net: /run/netns/wg
    capabilities:
     - CAP_NET_BIND_SERVICE
     - CAP_CHOWN
     - CAP_SETUID
     - CAP_SETGID
     - CAP_DAC_OVERRIDE

devices

To access the console, it's necessary to explicitly add a "device" definition, for example:

devices:
- path: "/dev/console"
  type: c
  major: 5
  minor: 1
  mode: 0666

See the getty package for a more complete example and see runc for context.

To grant access to all block devices use:

devices:
- path: all
  type: b

See the format package for an example.

Mount Options

When mounting filesystem paths into a container - whether as part of onboot or services - there are several options of which you need to be aware. Using them properly is necessary for your containers to function properly.

For most containers - e.g. nginx or even docker - these options are not needed. Simply doing the following will work fine:

binds:
 - /var:/some/var/path

Please note that binds doesn't add the mount points, but replaces them. You can examine the Dockerfile of the component (in particular, binds value of org.mobyproject.config label) to get the list of the existing binds.

However, in some circumstances you will need additional options. These options are used primarily if you intend to make changes to mount points from within your container that should be visible from outside the container, e.g., if you intend to mount an external disk from inside the container but have it be visible outside.

In order for new mounts from within a container to be propagated, you must set the following on the container:

  1. rootfsPropagation: shared
  2. The mount point into the container below which new mounts are to occur must be rshared,rbind. In practice, this is /var (or some subdir of /var), since that is the only true read-write area of the filesystem where you will mount things.

Thus, if you have a regular container that is only reading and writing, go ahead and do:

binds:
 - /var:/some/var/path

On the other hand, if you have a container that will make new mounts that you wish to be visible outside the container, do:

binds:
 - /var:/var:rshared,rbind
rootfsPropagation: shared