Fedora CoreOS (FCOS)

About FCOS

Fedora CoreOS (FCOS) represents the next generation of single-purpose container operating system technology by providing the quality standards of Fedora with automated, remote upgrade features.

FCOS is supported only as a component of OKD 4.12 for all OKD machines. FCOS is the only supported operating system for OKD control plane, or master, machines. While FCOS is the default operating system for all cluster machines, you can create compute machines, which are also known as worker machines, that use Fedora as their operating system. There are two general ways FCOS is deployed in OKD 4.12:

• If you install your cluster on infrastructure that the installation program provisions, FCOS images are downloaded to the target platform during installation. Suitable Ignition config files, which control the FCOS configuration, are also downloaded and used to deploy the machines.

• If you install your cluster on infrastructure that you manage, you must follow the installation documentation to obtain the FCOS images, generate Ignition config files, and use the Ignition config files to provision your machines.

Key FCOS features

The following list describes key features of the FCOS operating system:

• Based on Fedora: The underlying operating system consists primarily of Fedora components. The same quality, security, and control measures that support Fedora also support FCOS. For example, FCOS software is in RPM packages, and each FCOS system starts up with a Fedora kernel and a set of services that are managed by the systemd init system.

• Controlled immutability: Although it contains Fedora components, FCOS is designed to be managed more tightly than a default Fedora installation. Management is performed remotely from the OKD cluster. When you set up your FCOS machines, you can modify only a few system settings. This controlled immutability allows OKD to store the latest state of FCOS systems in the cluster so it is always able to create additional machines and perform updates based on the latest FCOS configurations.

• CRI-O container runtime: Although FCOS contains features for running the OCI- and libcontainer-formatted containers that Docker requires, it incorporates the CRI-O container engine instead of the Docker container engine. By focusing on features needed by Kubernetes platforms, such as OKD, CRI-O can offer specific compatibility with different Kubernetes versions. CRI-O also offers a smaller footprint and reduced attack surface than is possible with container engines that offer a larger feature set. At the moment, CRI-O is the only engine available within OKD clusters.

  CRI-O can use either the runC or crun container runtime to start and manage containers. For information about how to enable crun, see the documentation for creating a ContainerRuntimeConfig CR.

• Set of container tools: For tasks such as building, copying, and otherwise managing containers, FCOS replaces the Docker CLI tool with a compatible set of container tools. The podman CLI tool supports many container runtime features, such as running, starting, stopping, listing, and removing containers and container images. The skopeo CLI tool can copy, authenticate, and sign images. You can use the crictl CLI tool to work with containers and pods from the CRI-O container engine. While direct use of these tools in FCOS is discouraged, you can use them for debugging purposes.

• rpm-ostree upgrades: FCOS features transactional upgrades using the rpm-ostree system. Updates are delivered by means of container images and are part of the OKD update process. When deployed, the container image is pulled, extracted, and written to disk, then the bootloader is modified to boot into the new version. The machine will reboot into the update in a rolling manner to ensure cluster capacity is minimally impacted.

• bootupd firmware and bootloader updater: Package managers and hybrid systems such as rpm-ostree do not update the firmware or the bootloader. With bootupd, FCOS users have access to a cross-distribution, system-agnostic update tool that manages firmware and boot updates in UEFI and legacy BIOS boot modes that run on modern architectures, such as x86_64, ppc64le, and aarch64.

  For information about how to install bootupd, see the documentation for Updating the bootloader using bootupd.

• Updated through the Machine Config Operator: In OKD, the Machine Config Operator handles operating system upgrades. Instead of upgrading individual packages, as is done with yum upgrades, rpm-ostree delivers upgrades of the OS as an atomic unit. The new OS deployment is staged during upgrades and goes into effect on the next reboot. If something goes wrong with the upgrade, a single rollback and reboot returns the system to the previous state. FCOS upgrades in OKD are performed during cluster updates.

For FCOS systems, the layout of the rpm-ostree file system has the following characteristics:

• /usr is where the operating system binaries and libraries are stored and is read-only. We do not support altering this.

• /etc, /boot, /var are writable on the system but only intended to be altered by the Machine Config Operator.

• /var/lib/containers is the graph storage location for storing container images.

Choosing how to configure FCOS

FCOS is designed to deploy on an OKD cluster with a minimal amount of user configuration. In its most basic form, this consists of:

• Starting with a provisioned infrastructure, such as on AWS, or provisioning the infrastructure yourself.

• Supplying a few pieces of information, such as credentials and cluster name, in an install-config.yaml file when running openshift-install.

Because FCOS systems in OKD are designed to be fully managed from the OKD cluster after that, directly changing an FCOS machine is discouraged. Although limited direct access to FCOS machines cluster can be accomplished for debugging purposes, you should not directly configure FCOS systems. Instead, if you need to add or change features on your OKD nodes, consider making changes in the following ways:

• Kubernetes workload objects, such as DaemonSet and Deployment: If you need to add services or other user-level features to your cluster, consider adding them as Kubernetes workload objects. Keeping those features outside of specific node configurations is the best way to reduce the risk of breaking the cluster on subsequent upgrades.

• Day-2 customizations: If possible, bring up a cluster without making any customizations to cluster nodes and make necessary node changes after the cluster is up. Those changes are easier to track later and less likely to break updates. Creating machine configs or modifying Operator custom resources are ways of making these customizations.

• Day-1 customizations: For customizations that you must implement when the cluster first comes up, there are ways of modifying your cluster so changes are implemented on first boot. Day-1 customizations can be done through Ignition configs and manifest files during openshift-install or by adding boot options during ISO installs provisioned by the user.

Here are examples of customizations you could do on day 1:

• Kernel arguments: If particular kernel features or tuning is needed on nodes when the cluster first boots.

• Disk encryption: If your security needs require that the root file system on the nodes are encrypted, such as with FIPS support.

• Kernel modules: If a particular hardware device, such as a network card or video card, does not have a usable module available by default in the Linux kernel.

• Chronyd: If you want to provide specific clock settings to your nodes, such as the location of time servers.

To accomplish these tasks, you can augment the openshift-install process to include additional objects such as MachineConfig objects. Those procedures that result in creating machine configs can be passed to the Machine Config Operator after the cluster is up.

Choosing how to deploy FCOS

Differences between FCOS installations for OKD are based on whether you are deploying on an infrastructure provisioned by the installer or by the user:

• Installer-provisioned: Some cloud environments offer pre-configured infrastructures that allow you to bring up an OKD cluster with minimal configuration. For these types of installations, you can supply Ignition configs that place content on each node so it is there when the cluster first boots.

• User-provisioned: If you are provisioning your own infrastructure, you have more flexibility in how you add content to a FCOS node. For example, you could add kernel arguments when you boot the FCOS ISO installer to install each system. However, in most cases where configuration is required on the operating system itself, it is best to provide that configuration through an Ignition config.

The Ignition facility runs only when the FCOS system is first set up. After that, Ignition configs can be supplied later using the machine config.

About Ignition

Ignition is the utility that is used by FCOS to manipulate disks during initial configuration. It completes common disk tasks, including partitioning disks, formatting partitions, writing files, and configuring users. On first boot, Ignition reads its configuration from the installation media or the location that you specify and applies the configuration to the machines.

Whether you are installing your cluster or adding machines to it, Ignition always performs the initial configuration of the OKD cluster machines. Most of the actual system setup happens on each machine itself. For each machine, Ignition takes the FCOS image and boots the FCOS kernel. Options on the kernel command line identify the type of deployment and the location of the Ignition-enabled initial RAM disk (initramfs).

How Ignition works

To create machines by using Ignition, you need Ignition config files. The OKD installation program creates the Ignition config files that you need to deploy your cluster. These files are based on the information that you provide to the installation program directly or through an install-config.yaml file.

The way that Ignition configures machines is similar to how tools like cloud-init or Linux Anaconda kickstart configure systems, but with some important differences:

• Ignition runs from an initial RAM disk that is separate from the system you are installing to. Because of that, Ignition can repartition disks, set up file systems, and perform other changes to the machine’s permanent file system. In contrast, cloud-init runs as part of a machine init system when the system boots, so making foundational changes to things like disk partitions cannot be done as easily. With cloud-init, it is also difficult to reconfigure the boot process while you are in the middle of the node boot process.

• Ignition is meant to initialize systems, not change existing systems. After a machine initializes and the kernel is running from the installed system, the Machine Config Operator from the OKD cluster completes all future machine configuration.

• Instead of completing a defined set of actions, Ignition implements a declarative configuration. It checks that all partitions, files, services, and other items are in place before the new machine starts. It then makes the changes, like copying files to disk that are necessary for the new machine to meet the specified configuration.

• After Ignition finishes configuring a machine, the kernel keeps running but discards the initial RAM disk and pivots to the installed system on disk. All of the new system services and other features start without requiring a system reboot.

• Because Ignition confirms that all new machines meet the declared configuration, you cannot have a partially configured machine. If a machine setup fails, the initialization process does not finish, and Ignition does not start the new machine. Your cluster will never contain partially configured machines. If Ignition cannot complete, the machine is not added to the cluster. You must add a new machine instead. This behavior prevents the difficult case of debugging a machine when the results of a failed configuration task are not known until something that depended on it fails at a later date.

• If there is a problem with an Ignition config that causes the setup of a machine to fail, Ignition will not try to use the same config to set up another machine. For example, a failure could result from an Ignition config made up of a parent and child config that both want to create the same file. A failure in such a case would prevent that Ignition config from being used again to set up an other machines until the problem is resolved.

• If you have multiple Ignition config files, you get a union of that set of configs. Because Ignition is declarative, conflicts between the configs could cause Ignition to fail to set up the machine. The order of information in those files does not matter. Ignition will sort and implement each setting in ways that make the most sense. For example, if a file needs a directory several levels deep, if another file needs a directory along that path, the later file is created first. Ignition sorts and creates all files, directories, and links by depth.

• Because Ignition can start with a completely empty hard disk, it can do something cloud-init cannot do: set up systems on bare metal from scratch using features such as PXE boot. In the bare metal case, the Ignition config is injected into the boot partition so that Ignition can find it and configure the system correctly.

The Ignition sequence

The Ignition process for an FCOS machine in an OKD cluster involves the following steps:

• The machine gets its Ignition config file. Control plane machines get their Ignition config files from the bootstrap machine, and worker machines get Ignition config files from a control plane machine.

• Ignition creates disk partitions, file systems, directories, and links on the machine. It supports RAID arrays but does not support LVM volumes.

• Ignition mounts the root of the permanent file system to the /sysroot directory in the initramfs and starts working in that /sysroot directory.

• Ignition configures all defined file systems and sets them up to mount appropriately at runtime.

• Ignition runs systemd temporary files to populate required files in the /var directory.

• Ignition runs the Ignition config files to set up users, systemd unit files, and other configuration files.

• Ignition unmounts all components in the permanent system that were mounted in the initramfs.

• Ignition starts up the init process of the new machine, which in turn starts up all other services on the machine that run during system boot.

At the end of this process, the machine is ready to join the cluster and does not require a reboot.

Viewing Ignition configuration files

To see the Ignition config file used to deploy the bootstrap machine, run the following command:

$ openshift-install create ignition-configs --dir $HOME/testconfig

After you answer a few questions, the bootstrap.ign, master.ign, and worker.ign files appear in the directory you entered.

To see the contents of the bootstrap.ign file, pipe it through the jq filter. Here’s a snippet from that file:

$ cat $HOME/testconfig/bootstrap.ign | jq
{
  "ignition": {
    "version": "3.2.0"
  },
  "passwd": {
    "users": [
      {
        "name": "core",
        "sshAuthorizedKeys": [
          "ssh-rsa AAAAB3NzaC1yc...."
        ]
      }
    ]
  },
  "storage": {
    "files": [
      {
        "overwrite": false,
        "path": "/etc/motd",
        "user": {
          "name": "root"
        },
        "append": [
          {
            "source": "data:text/plain;charset=utf-8;base64,VGhpcyBpcyB0aGUgYm9vdHN0cmFwIG5vZGU7IGl0IHdpbGwgYmUgZGVzdHJveWVkIHdoZW4gdGhlIG1hc3RlciBpcyBmdWxseSB1cC4KClRoZSBwcmltYXJ5IHNlcnZpY2VzIGFyZSByZWxlYXNlLWltYWdlLnNlcnZpY2UgZm9sbG93ZWQgYnkgYm9vdGt1YmUuc2VydmljZS4gVG8gd2F0Y2ggdGhlaXIgc3RhdHVzLCBydW4gZS5nLgoKICBqb3VybmFsY3RsIC1iIC1mIC11IHJlbGVhc2UtaW1hZ2Uuc2VydmljZSAtdSBib290a3ViZS5zZXJ2aWNlCg=="
          }
        ],
        "mode": 420
      },
...

To decode the contents of a file listed in the bootstrap.ign file, pipe the base64-encoded data string representing the contents of that file to the base64 -d command. Here’s an example using the contents of the /etc/motd file added to the bootstrap machine from the output shown above:

$ echo VGhpcyBpcyB0aGUgYm9vdHN0cmFwIG5vZGU7IGl0IHdpbGwgYmUgZGVzdHJveWVkIHdoZW4gdGhlIG1hc3RlciBpcyBmdWxseSB1cC4KClRoZSBwcmltYXJ5IHNlcnZpY2VzIGFyZSByZWxlYXNlLWltYWdlLnNlcnZpY2UgZm9sbG93ZWQgYnkgYm9vdGt1YmUuc2VydmljZS4gVG8gd2F0Y2ggdGhlaXIgc3RhdHVzLCBydW4gZS5nLgoKICBqb3VybmFsY3RsIC1iIC1mIC11IHJlbGVhc2UtaW1hZ2Uuc2VydmljZSAtdSBib290a3ViZS5zZXJ2aWNlCg== | base64 --decode

Example output

This is the bootstrap node; it will be destroyed when the master is fully up.
The primary services are release-image.service followed by bootkube.service. To watch their status, run e.g.
  journalctl -b -f -u release-image.service -u bootkube.service

Repeat those commands on the master.ign and worker.ign files to see the source of Ignition config files for each of those machine types. You should see a line like the following for the worker.ign, identifying how it gets its Ignition config from the bootstrap machine:

"source": "https://api.myign.develcluster.example.com:22623/config/worker",

Here are a few things you can learn from the bootstrap.ign file:

• Format: The format of the file is defined in the Ignition config spec. Files of the same format are used later by the MCO to merge changes into a machine’s configuration.

• Contents: Because the bootstrap machine serves the Ignition configs for other machines, both master and worker machine Ignition config information is stored in the bootstrap.ign, along with the bootstrap machine’s configuration.

• Size: The file is more than 1300 lines long, with path to various types of resources.

• The content of each file that will be copied to the machine is actually encoded into data URLs, which tends to make the content a bit clumsy to read. (Use the jq and base64 commands shown previously to make the content more readable.)

• Configuration: The different sections of the Ignition config file are generally meant to contain files that are just dropped into a machine’s file system, rather than commands to modify existing files. For example, instead of having a section on NFS that configures that service, you would just add an NFS configuration file, which would then be started by the init process when the system comes up.

• users: A user named core is created, with your SSH key assigned to that user. This allows you to log in to the cluster with that user name and your credentials.

• storage: The storage section identifies files that are added to each machine. A few notable files include /root/.docker/config.json (which provides credentials your cluster needs to pull from container image registries) and a bunch of manifest files in /opt/openshift/manifests that are used to configure your cluster.

• systemd: The systemd section holds content used to create systemd unit files. Those files are used to start up services at boot time, as well as manage those services on running systems.

• Primitives: Ignition also exposes low-level primitives that other tools can build on.

Changing Ignition configs after installation

Machine config pools manage a cluster of nodes and their corresponding machine configs. Machine configs contain configuration information for a cluster. To list all machine config pools that are known:

$ oc get machineconfigpools

Example output

NAME   CONFIG                                  UPDATED UPDATING DEGRADED
master master-1638c1aea398413bb918e76632f20799 False   False    False
worker worker-2feef4f8288936489a5a832ca8efe953 False   False    False

To list all machine configs:

$ oc get machineconfig

Example output

NAME                                      GENERATEDBYCONTROLLER   IGNITIONVERSION   CREATED   OSIMAGEURL
00-master                                 4.0.0-0.150.0.0-dirty   3.2.0             16m
00-master-ssh                             4.0.0-0.150.0.0-dirty                     16m
00-worker                                 4.0.0-0.150.0.0-dirty   3.2.0             16m
00-worker-ssh                             4.0.0-0.150.0.0-dirty                     16m
01-master-kubelet                         4.0.0-0.150.0.0-dirty   3.2.0             16m
01-worker-kubelet                         4.0.0-0.150.0.0-dirty   3.2.0             16m
master-1638c1aea398413bb918e76632f20799   4.0.0-0.150.0.0-dirty   3.2.0             16m
worker-2feef4f8288936489a5a832ca8efe953   4.0.0-0.150.0.0-dirty   3.2.0             16m

The Machine Config Operator acts somewhat differently than Ignition when it comes to applying these machine configs. The machine configs are read in order (from 00* to 99*). Labels inside the machine configs identify the type of node each is for (master or worker). If the same file appears in multiple machine config files, the last one wins. So, for example, any file that appears in a 99* file would replace the same file that appeared in a 00* file. The input MachineConfig objects are unioned into a “rendered” MachineConfig object, which will be used as a target by the operator and is the value you can see in the machine config pool.

To see what files are being managed from a machine config, look for “Path:” inside a particular MachineConfig object. For example:

$ oc describe machineconfigs 01-worker-container-runtime | grep Path:

Example output

            Path:            /etc/containers/registries.conf
            Path:            /etc/containers/storage.conf
            Path:            /etc/crio/crio.conf

Be sure to give the machine config file a later name (such as 10-worker-container-runtime). Keep in mind that the content of each file is in URL-style data. Then apply the new machine config to the cluster.