I’ve been operating our small company’s kubernetes cluster for more or less two years by now. I usually construct the cluster using Rancher Kubernetes Engine in Hetzner Cloud. Hetzner itself doesn’t have managed Kubernetes offering, so we assembled nodes by ourselves using the Hetzner Cloud instances.

In some rare cases, we need to access the nodes. We can do that with some of the options here:

• Old fashioned public key ssh login
• Hetzner CLI hcloud server ssh command
• Rancher CLI rancher ssh command
• Using Kubernetes pods via Kubernetes controlplane or kubectl access

I’ve been preparing myself to get CKA (Certified Kubernetes Administrator) and using Kodekloud as the lab/course provider. In a specific section, the lab has task puzzle. The lab prepares a Static Pods in one of the node, then we are tasked to delete that static pods permanently.

In case you are wondering, Static Pods is a mechanism available for Kubelet to spawn pods by themselves without using Kubernetes controlplane. You specify Pod manifests and store it in a specific directory in the node. Kubelet will watch the manifests directory and deploy the pods, and even recreate it if the spec changes. This provides a way for Kubelet agent to bootstrap Kubernetes controlplane or main component before the controlplane even exists.

Long story short, the lab provides access only to the controlplane node. The static pods is in a node called node01. If you want to delete the pods, you need to get inside node01 and then delete the static pods manifests yaml file yourself. You can’t delete the pods via controlplane because kubelet agent in node01 will recreate the pods again as long as the pod spec is inside the manifests directory.

So, how we do get inside the node?

Well, obvious answer is via SSH. The lab conveniently prepared such that controlplane public key are stored in node01. That means you can just kubectl get nodes -o wide and get the IP address of node01 then SSH there.

But… What if the node was prepared via cloud provider or third party in such a way that controlplane doesn’t have direct SSH access to these nodes? For example, when we construct k3s cluster, the node only needs controlplane address and join token. Controlplane doesn’t have to access the nodes via SSH, which is safer. This way, bootstrap script can run without having to be run from inside controlplane. I imagine it’s useful to create new clusters via GitOps, like terraform.

If you have Cluster Admin role, or at least able to modify Pod security policies, then you can access the nodes using host namespace. Note that namespace in this context is the Linux kernel namespace and not Kubernetes namespace resource object.

Using nsenter command

The logic are fairly simple. Use Kubernetes controlplane/api-server to create a pod that shares namespace with the root process in the node itself. Then we use kubectl exec to get access to the shell inside the container. The shell then acts as if it was the root process of the node.

Create a pod or deployment spec that uses nsenter

apiVersion: v1
kind: Pod
metadata:
  name: root-shell
  namespace: kube-system
spec:
  containers:
    - name: shell
      image: alpine
      command:
        - nsenter
      args:
        - '-t'
        - '1'
        - '-m'
        - '-u'
        - '-i'
        - '-n'
        - tail
        - '-f'
        - /dev/null
      securityContext:
        privileged: true
  hostNetwork: true
  hostPID: true
  hostIPC: true
  nodeName: node01

These are some explanations of what we are trying to do.

From the key metadata it means, we create a pod with a name root-shell (names are arbitrary). We create it in kube-system k8s namespace, because usually only cluster-admin are allowed to access this namespace.

From the key spec.containers[] we create a container with name shell (names are arbitrary too). We uses image: alpine because it is lightweight and quick to pull. We then execute program nsenter, which is a program to execute another program using the specified Linux namespace.

The arguments used are fairly straightforward:

We use PID 1, which is a root PID using argument -t 1. We use argument -m -u -i -n which basically means: recognize resources owned by the target PID such as all the mount (for file system and device access), UTS (for host identifier/identity), IPC (resources in /proc), and network (interfaces, routes, ip command, etc).

The next argument is just the program we want to run. In this case, I just use tail -f /dev/null so that the process hangs and stays indefinitely without being terminated, so that the pod stays up.

The other keys are used to allow security policies/features.

The key spec.containers[].securityContext.privileged: true is specified so we can execute nsenter with target PID 1, which is basically a privilege escalation. The keys spec.hostNetwork, spec.hostPID, spec.hostIPC are used so that the Pod Admission Controller can check these pod spec request against the role the Kubernetes user is allowed to. It is set to false by default, so the pod will not get scheduled to the node. We set it to true now. Also this is assuming that your kubectl user role does have this kind of permission.

Lastly, the key spec.nodeName is just to specify where you want this pod to be placed. If you want to access node01 then you use this node name.

Once the pod is running in the node, you can proceed to the next step.

Use kubectl exec to attach shell

Simply uses kubectl exec -it -n kube-system root-shell -- sh to attach to the shell. You now gain access to root process. Any command you execute in this shell is as if it was executed by root themselves, so you need to be careful.

Wrapping up

After you are doing what you must, for example ad-hoc security patches, installing packages, etc, do not forget to delete your pods. Leaving this kind of pod running is a security hole. In some cases, it is probably better to use sleep command as some kind of session timer. For example if you use nsenter like this:

nsenter -t 1 -m -u -i -n sleep 3600

This means the pod will stays up for 3600 seconds (an hour). Then you can make sure that the pod will not restart once the time runs out by specifying spec.restartPolicy: Never in the pod spec.