Enable seccomp for all workloads with a new v1.22 alpha feature

Wednesday, August 25, 2021

Author: Sascha Grunert, Red Hat

This blog post is about a new Kubernetes feature introduced in v1.22, which adds an additional security layer on top of the existing seccomp support. Seccomp is a security mechanism for Linux processes to filter system calls (syscalls) based on a set of defined rules. Applying seccomp profiles to containerized workloads is one of the key tasks when it comes to enhancing the security of the application deployment. Developers, site reliability engineers and infrastructure administrators have to work hand in hand to create, distribute and maintain the profiles over the applications life-cycle.

You can use the securityContext field of Pods and their containers can be used to adjust security related configurations of the workload. Kubernetes introduced dedicated seccomp related API fields in this SecurityContext with the in v1.19.0. This enhancement allowed an easier way to specify if the whole pod or a specific container should run as:

Unconfined: seccomp will not be enabled
RuntimeDefault: the container runtimes default profile will be used
Localhost: a node local profile will be applied, which is being referenced by a relative path to the seccomp profile root (<kubelet-root-dir>/seccomp) of the kubelet

With the graduation of seccomp, nothing has changed from an overall security perspective, because Unconfined is still the default. This is totally fine if you consider this from the upgrade path and backwards compatibility perspective of Kubernetes releases. But it also means that it is more likely that a workload runs without seccomp at all, which should be fixed in the long term.

`SeccompDefault` to the rescue

Kubernetes v1.22.0 introduces a new kubelet feature gate SeccompDefault, which has been added in alpha state as every other new feature. This means that it is disabled by default and can be enabled manually for every single Kubernetes node.

What does the feature do? Well, it just changes the default seccomp profile from Unconfined to RuntimeDefault. If not specified differently in the pod manifest, then the feature will add a higher set of security constraints by using the default profile of the container runtime. These profiles may differ between runtimes like

Enabling the feature

Two kubelet configuration changes have to be made to enable the feature:

Enable the feature gate by setting the SeccompDefault=true via the command line (--feature-gates) or the kubelet configuration file.
Turn on the feature by enabling the feature by adding the --seccomp-default command line flag or via the kubelet configuration file (seccompDefault: true).

The kubelet will error on startup if only one of the above steps have been done.

Trying it out

If the feature is enabled on a node, then you can create a new workload like this:

apiVersion: v1
kind: Pod
metadata:
  name: test-pod
spec:
  containers:
    - name: test-container
      image: nginx:1.21

Now it is possible to inspect the used seccomp profile by using runtime specification:

CONTAINER_ID=$(sudo crictl ps -q --name=test-container)
sudo crictl inspect $CONTAINER_ID | jq .info.runtimeSpec.linux.seccomp

{
  "defaultAction": "SCMP_ACT_ERRNO",
  "architectures": ["SCMP_ARCH_X86_64", "SCMP_ARCH_X86", "SCMP_ARCH_X32"],
  "syscalls": [
    {
      "names": ["_llseek", "_newselect", "accept", …, "write", "writev"],
      "action": "SCMP_ACT_ALLOW"
    },
    …
  ]
}

You can see that the lower level container runtime (

Please note that the feature will not influence any Kubernetes API for now. Therefore, it is not possible to retrieve the used seccomp profile via kubectl get or describe if the SecurityContext.

The feature also works when using multiple containers within a pod, for example if you create a pod like this:

apiVersion: v1
kind: Pod
metadata:
  name: test-pod
spec:
  containers:
    - name: test-container-nginx
      image: nginx:1.21
      securityContext:
        seccompProfile:
          type: Unconfined
    - name: test-container-redis
      image: redis:6.2

then you should see that the test-container-nginx runs without a seccomp profile:

sudo crictl inspect $(sudo crictl ps -q --name=test-container-nginx) |
    jq '.info.runtimeSpec.linux.seccomp == null'
true

Whereas the container test-container-redis runs with RuntimeDefault:

sudo crictl inspect $(sudo crictl ps -q --name=test-container-redis) |
    jq '.info.runtimeSpec.linux.seccomp != null'
true

The same applies to the pod itself, which also runs with the default profile:

sudo crictl inspectp (sudo crictl pods -q --name test-pod) |
    jq '.info.runtimeSpec.linux.seccomp != null'
true

Upgrade strategy

It is recommended to enable the feature in multiple steps, whereas different risks and mitigations exist for each one.

Feature gate enabling

Enabling the feature gate at the kubelet level will not turn on the feature, but will make it possible by using the SeccompDefault kubelet configuration or the --seccomp-default CLI flag. This can be done by an administrator for the whole cluster or only a set of nodes.

Testing the Application

If you're trying this within a dedicated test environment, you have to ensure that the application code does not trigger syscalls blocked by the RuntimeDefault profile before enabling the feature on a node. This can be done by:

Recommended: Analyzing the code (manually or by running the application with Unconfined. Alternatively, you can create a custom seccomp profile (see optional step below). profile based on the default by adding the additional syscalls to the "action": "SCMP_ACT_ALLOW" section.

Recommended: Manually set the profile to the target workload and use a rolling upgrade to deploy into production. Rollback the deployment if the application does not work as intended.

Optional: Run the application against an end-to-end test suite to trigger all relevant code paths with RuntimeDefault enabled. If a test fails, use the same mitigation as mentioned above.

Optional: Create a custom seccomp profile based on the default and change its default action from SCMP_ACT_ERRNO to SCMP_ACT_LOG. This means that the seccomp filter for unknown syscalls will have no effect on the application at all, but the system logs will now indicate which syscalls may be blocked. This requires at least a Kernel version 4.14 as well as a recent release. Monitor the application hosts audit logs (defaults to /var/log/audit/audit.log) or syslog entries (defaults to /var/log/syslog) for syscalls via type=SECCOMP (for audit) or type=1326 (for syslog). Compare the syscall ID with those and add them to the custom profile. Be aware that custom audit policies may lead into missing syscalls, depending on the configuration of auditd.

Optional: Use cluster additions like the for profiling the application via its

Deploying the modified application

Based on the outcome of the application tests, it may be required to change the application deployment by either specifying Unconfined or a custom seccomp profile. This is not the case if the application works as intended with RuntimeDefault.

Enable the kubelet configuration

If everything went well, then the feature is ready to be enabled by the kubelet configuration or its corresponding CLI flag. This should be done on a per-node basis to reduce the overall risk of missing a syscall during the investigations when running the application tests. If it's possible to monitor audit logs within the cluster, then it's recommended to do this for eventually missed seccomp events. If the application works as intended then the feature can be enabled for further nodes within the cluster.

Conclusion

Thank you for reading this blog post! I hope you enjoyed to see how the usage of seccomp profiles has been evolved in Kubernetes over the past releases as much as I do. On your own cluster, change the default seccomp profile to RuntimeDefault (using this new feature) and see the security benefits, and, of course, feel free to reach out any time for feedback or questions.

Editor's note: If you have any questions or feedback about this blog post, feel free to reach out via the .

←Previous