Reboot strategies on updates

    The overarching goal of Flatcar Container Linux is to secure the Internet’s backend infrastructure. We believe that automatically updating the operating system is one of the best tools to achieve this goal.

    We realize that each Flatcar Container Linux cluster has a unique tolerance for risk and the operational needs of your applications are complex. In order to meet everyone’s needs, there are different update/reboot strategies that we have developed.

    This document is about the update client and how it consumes the updates when they get available. The public update server makes the new releases available as soon as they get published. To control this part of the update rollout, look at the different public update channels and how you can run your own update server .

    It’s important to note that updates are always downloaded to the passive partition when they become available (see further below for disabling automatic updates). A reboot is the last step of the update, where the active and passive partitions are swapped ( rollback instructions ).

    The reboot is done by the reboot manager, by default this is the locksmithd.service included on the image. For Kubernetes the recommended reboot manager is FLUO which replaces locksmithd because it knows how to gracefully reboot a Kubernetes node. The kured reboot manager will be supported as well starting from Flatcar versions with a release number greater than 3067.0.0.

    The update-engine.service responsible for downloading and applying the updates can be in different states which you can query with update_engine_client -status:

    • UPDATE_STATUS_IDLE (did not find an update)
    • UPDATE_STATUS_CHECKING_FOR_UPDATE
    • UPDATE_STATUS_UPDATE_AVAILABLE (can be a result of update_engine_client -check_for_update)
    • UPDATE_STATUS_DOWNLOADING
    • UPDATE_STATUS_VERIFYING
    • UPDATE_STATUS_FINALIZING
    • UPDATE_STATUS_UPDATED_NEED_REBOOT (update applied to inactive partition, this is where the reboot manager comes in)
    • UPDATE_STATUS_REPORTING_ERROR_EVENT (error encountered, use journalctl -u update-engine -e to get more info)

    Locksmithd reboot strategies

    These locksmithd strategies control how a reboot occurs when update-engine indicates that one is needed:

    Strategy Description
    etcd-lock Reboot after first taking a distributed lock in etcd (reboot window applies)
    reboot Reboot immediately after an update is applied (reboot window applies)
    off Do not reboot after updates are applied

    You can configure a reboot window in which reboots are allowed to happen through one of the strategies.

    The default behavior is reboot and results in a reboot with a delay of 5 minutes.

    Reboot strategy options through CLC/Ignition

    The reboot strategy can be set with a special Container Linux Config (CLC) section:

    locksmith:
      reboot_strategy: "etcd-lock"
    

    This gets transpiled to the following Ignition configuration which writes the line REBOOT_STRATEGY="etcd-lock" to /etc/flatcar/update.conf:

    {
      "ignition": {
        "version": "2.3.0"
      },
      "storage": {
        "files": [
          {
            "filesystem": "root",
            "path": "/etc/flatcar/update.conf",
            "contents": {
              "source": "data:,%0AREBOOT_STRATEGY%3D%22etcd-lock%22",
              "verification": {}
            },
            "mode": 420
          }
        ]
      }
    }
    

    Note: You must not combine the special locksmith: CLC section with a CLC files: section that writes to the /etc/flatcar/update.conf file (or /etc/coreos/update.conf through the symlinked /etc/coreos folder). This would result in a conflict where only one entry wins.

    etcd-lock

    The etcd-lock strategy mandates that each machine acquire and hold a reboot lock before it is allowed to reboot. The main goal behind this strategy is to allow for an update to be applied to a cluster quickly, without losing the quorum membership in etcd or rapidly reducing capacity for the services running on the cluster. The reboot lock is held until the machine releases it after a successful update.

    The number of machines allowed to reboot simultaneously is configurable via a command line utility:

    $ locksmithctl set-max 4
    Old: 1
    New: 4
    

    This setting is stored in etcd so it won’t have to be configured for subsequent machines.

    To view the number of available slots and find out which machines in the cluster are holding locks, run:

    $ locksmithctl status
    Available: 0
    Max: 1
    
    MACHINE ID
    69d27b356a94476da859461d3a3bc6fd
    

    If needed, you can manually clear a lock by providing the machine ID:

    locksmithctl unlock 69d27b356a94476da859461d3a3bc6fd
    

    Reboot immediately

    The reboot strategy works exactly like it sounds: the machine is rebooted as soon as the update has been installed to the passive partition. If the applications running on your cluster are highly resilient, this strategy was made for you.

    Off

    The off strategy is also straightforward. The update will be installed onto the passive partition and await a reboot command to complete the update. We don’t recommend this strategy unless you reboot frequently as part of your normal operations workflow.

    Read below on how to disable automatic updates if this is what you actually want to achieve instead of having a half applied update on disk that gets selected even on an accidental reboot. It is also blocking the inactive partition with the earliest version that gets available as update, requiring the double update workaround at the end of this document.

    Auto-updates with a maintenance window

    Locksmith supports maintenance windows in addition to the reboot strategies mentioned earlier. Maintenance windows define a window of time during which a reboot can occur. These operate in addition to reboot strategies, so if the machine has a maintenance window and requires a reboot lock, the machine will only reboot when it has the lock during that window.

    Windows are defined by a start time and a length. In this example, the window is defined to be every Thursday between 04:00 and 05:00:

    locksmith:
      reboot_strategy: reboot
      window_start: Thu 04:00
      window_length: 1h
    

    This will configure a Flatcar Container Linux machine to follow the reboot strategy, and thus when an update is ready it will simply reboot instead of attempting to grab a lock in etcd. This machine however has also been configured to only reboot between 04:00 and 05:00 on Thursdays, so if an update occurs outside of this window the machine will then wait until it is inside of this window to reboot.

    For more information about the supported syntax, refer to the Locksmith documentation .

    Updating PXE/iPXE machines

    PXE/iPXE machines download a new copy of Flatcar Container Linux every time they are started thus are dependent on the version of Flatcar Container Linux they are served. If you don’t automatically load new Flatcar Container Linux images into your PXE/iPXE server, your machines will never have new features or security updates.

    An easy solution to this problem is to use iPXE and reference images directly from the Flatcar Container Linux storage site . The alpha URL is automatically pointed to the new version of Flatcar Container Linux as it is released.

    Disable Automatic Updates

    If for a short time frame you want to temporarily disable update reboots, run sudo systemctl stop update-engine locksmithd, and when done, sudo systemctl start update-engine locksmithd.

    In case when you want to permanently disable automatic updates, it’s not recommended to mask the services because it makes it harder to manually apply updates. It’s rather recommended to overwrite the SERVER variable in the update configuration to an invalid value.

    You can configure this with a Container Linux Config (needs to be transpiled to Ignition JSON):

    storage:
      files:
        - path: /etc/flatcar/update.conf
          mode: 0644
          contents:
            inline: |
                        SERVER=disabled
    

    To manually run updates, remove the file and run update_engine_client -update or wait for the update to happen. After update-engine applied the update to the passive partition, you can already create the file again to disable automatic updates.

    The flatcar-update tool automatically removes the SERVER=disabled line to apply a manual update and restores it after applying the update (it also has an explicit --disable-afterwards switch to set SERVER=disabled):

    $ # For example, update to the latest Stable release:
    $ VER=$(curl -fsSL https://stable.release.flatcar-linux.net/amd64-usr/current/version.txt | grep FLATCAR_VERSION= | cut -d = -f 2)
    $ sudo flatcar-update --to-version $VER
    

    In case you didn’t have SERVER=disabled set you should use the --disable-afterwards switch to set SERVER=disabled in /etc/flatcar/update.conf which disables updates. Disabling updates ensures that you will stay on the version you specified.

    After applying the update, wait for the reboot to happen or invoke it manually.

    As alternative you could mask the update-engine and locksmithd services as follows (but read the warning below):

    systemd:
      units:
        - name: update-engine.service
          mask: true
        - name: locksmithd.service
          mask: true
    

    Note: As said, it’s not recommended to mask the services but if you want to manually trigger an update after having masked update-engine, you’ll need to unmask the service, start update-engine to trigger an update, and keep the service unmasked until the next reboot is completed and update-engine started and marked the updated partition as successful. Otherwise, the update will be considered unsuccessful and in all following reboots GRUB will use the old partition again because update-engine never marked the new partition to be successfully booted.

    To check that you can stop and mask update-engine after the reboot, run these commands to see that the partition was marked as successful. This will happen after the service ran for about 1 minute:

    $ sudo cgpt show "$(rootdev -s /usr)" | grep successful=1
                                      Attr: priority=1 tries=0 successful=1
    

    Airgapped updates

    Updating a machine without Internet access is done in two steps. First, you need to download the update payload on a non-airgapped machine, then you copy it to your airgapped machine and run the flatcar-update tool. If the flatcar-update tool is missing on your machine, download it first, too.

    On the non-airgapped machine (here for amd64, use ARCH=arm64 for arm64):

    ARCH=amd64
    VER=$(curl -fsSL https://stable.release.flatcar-linux.net/${ARCH}-usr/current/version.txt | grep FLATCAR_VERSION= | cut -d = -f 2)
    echo "$VER"
    # or if you know which version to update to, set it like VER=3033.2.1 (no channel info needed)
    wget "https://update.release.flatcar-linux.net/${ARCH}-usr/${VER}/flatcar_production_update.gz"
    

    On the airgapped machine (here with the file flatcar_production_update.gz in the current folder):

    VER=... # use the same value as above
    sudo ./flatcar-update --to-version "$VER" --to-payload flatcar_production_update.gz
    

    Then reboot or wait for the reboot coordinator to do so.

    Updating behind a proxy

    Public Internet access is required to contact CoreUpdate and download new versions of Flatcar Container Linux. If direct access is not available the update-engine service may be configured to use a HTTP or SOCKS proxy using curl-compatible environment variables, such as HTTPS_PROXY or ALL_PROXY. See curl’s documentation for details.

    systemd:
      units:
        - name: update-engine.service
          dropins:
            - name: 50-proxy.conf
              contents: |
                [Service]
                Environment=ALL_PROXY=http://proxy.example.com:3128            
    

    Proxy environment variables can also be set system-wide .

    Manually triggering an update

    Each machine should check in about 10 minutes after boot and roughly every hour after that. If you’d like to see it sooner, you can force an update check, which will skip any rate-limiting settings that are configured in CoreUpdate.

    $ update_engine_client -check_for_update
    [0123/220706:INFO:update_engine_client.cc(245)] Initiating update check and install.
    

    Double update workaround

    If you have disabled automatic reboots, and your host has already applied an update then your flatcar host will not apply a newer update until it has rebooted into the prior-applied update. ( i.e. Host is in UPDATE_STATUS_UPDATED_NEED_REBOOT state). To work around this intermediate reboot, one can call:

    update_engine_client -reset_status
    update_engine_client -check_for_update
    

    Configure a post-install update hook

    Sometimes you may want to run a custom action after update-engine wrote the new partition out. You can create a /usr/share/oem/bin/oem-postinst script that gets two arguments passed. The first argument is the slot (A or B), the second is the temporary mount point where the new /usr partition contents can be accessed. Since the hook runs shortly before the new partition is prioritized, you should not directly reboot there. Also you should only let the script exit with an error return code if you want to stop the update. The hook runs as root user process under the update-engine.service unit.

    The following CLC example shows how a custom reboot hook for kured can be added to old Flatcar releases that don’t support it yet (for release number greater than 3067.0.0 this is not needed).

    storage:
      filesystems:
        - name: oem
          mount:
            device: /dev/disk/by-label/OEM
            format: btrfs
            label: OEM
      directories:
      - path: /bin
        filesystem: oem
        mode: 0755
      files:
      - path: /bin/oem-postinst
        filesystem: oem
        mode: 0755
        contents:
          inline: |
            #!/bin/sh
            touch /run/reboot-required