Provide IAM credentials to containers running inside a kubernetes cluster based on annotations.

Traditionally in AWS, service level isolation is done using IAM roles. IAM roles are attributed through instance profiles and are accessible by services through the transparent usage by the aws-sdk of the ec2 metadata API. When using the aws-sdk, a call is made to the EC2 metadata API which provides temporary credentials that are then used to make calls to the AWS service.

The problem is that in a multi-tenanted containers based world, multiple containers will be sharing the underlying nodes. Given containers will share the same underlying nodes, providing access to AWS resources via IAM roles would mean that one needs to create an IAM role which is a union of all IAM roles. This is not acceptable from a security perspective.

The solution is to redirect the traffic that is going to the ec2 metadata API for docker containers to a container running on each instance, make a call to the AWS API to retrieve temporary credentials and return these to the caller. Other calls will be proxied to the EC2 metadata API. This container will need to run with host networking enabled so that it can call the EC2 metadata API itself.

It is necessary to create an IAM role which can assume other roles and assign it to each kubernetes worker and list regions.
List regions required permissions because aws-go-sdk-v2 doesn't include regions list.

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Action": [
        "sts:AssumeRole"
      ],
      "Effect": "Allow",
      "Resource": "*"
    },
    {
      "Action": [
        "ec2:DescribeRegions"
      ],
      "Effect": "Allow",
      "Resource": "*"
    },
  ]
}

The roles that will be assumed must have a Trust Relationship which allows them to be assumed by the kubernetes worker role. See this StackOverflow post for more details.

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Sid": "",
      "Effect": "Allow",
      "Principal": {
        "Service": "ec2.amazonaws.com"
      },
      "Action": "sts:AssumeRole"
    },
    {
      "Sid": "",
      "Effect": "Allow",
      "Principal": {
        "AWS": "arn:aws:iam::123456789012:role/kubernetes-worker-role"
      },
      "Action": "sts:AssumeRole"
    }
  ]
}

Run the kube2iam container as a daemonset (so that it runs on each worker) with hostNetwork: true. The kube2iam daemon and iptables rule (see below) need to run before all other pods that would require access to AWS resources.

apiVersion: apps/v1
kind: DaemonSet
metadata:
  name: kube2iam
  labels:
    app: kube2iam
spec:
  selector:
    matchLabels:
      name: kube2iam
  template:
    metadata:
      labels:
        name: kube2iam
    spec:
      hostNetwork: true
      containers:
        - image: jtblin/kube2iam:latest
          name: kube2iam
          args:
            - "--base-role-arn=arn:aws:iam::123456789012:role/"
            - "--node=$(NODE_NAME)"
          env:
            - name: NODE_NAME
              valueFrom:
                fieldRef:
                  fieldPath: spec.nodeName
          ports:
            - containerPort: 8181
              hostPort: 8181
              name: http

To prevent containers from directly accessing the EC2 metadata API and gaining unwanted access to AWS resources, the traffic to 169.254.169.254 must be proxied for docker containers.

iptables \
  --append PREROUTING \
  --protocol tcp \
  --destination 169.254.169.254 \
  --dport 80 \
  --in-interface docker0 \
  --jump DNAT \
  --table nat \
  --to-destination `curl 169.254.169.254/latest/meta-data/local-ipv4`:8181

This rule can be added automatically by setting --iptables=true, setting the HOST_IP environment variable, and running the container in a privileged security context.

Warning: It is possible that other pods are started on an instance before kube2iam has started. Using --iptables=true (instead of applying the rule before starting the kubelet) could give those pods the opportunity to access the real EC2 metadata API, assume the role of the EC2 instance and thereby have all permissions the instance role has (including assuming potential other roles). Use with care if you don't trust the users of your kubernetes cluster or if you are running pods (that could be exploited) that have permissions to create other pods (e.g. controllers / operators).

Note that the interface --in-interface above or using the --host-interface cli flag may be different than docker0 depending on which virtual network you use e.g.

• for Calico, use cali+ (the interface name is something like cali1234567890)
• for kops (on kubenet), use cbr0
• for CNI, use cni0
• for EKS/amazon-vpc-cni-k8s, even with calico installed uses eni+. (Each pod gets an interface like eni4c0e15dfb05)
• for weave use weave
• for flannel use cni0
• for kube-router use kube-bridge
• for OpenShift use tun0
• for Cilium use lxc+

apiVersion: apps/v1
kind: DaemonSet
metadata:
  name: kube2iam
  labels:
    app: kube2iam
spec:
  selector:
    matchLabels:
      name: kube2iam
  template:
    metadata:
      labels:
        name: kube2iam
    spec:
      hostNetwork: true
      containers:
        - image: jtblin/kube2iam:latest
          name: kube2iam
          args:
            - "--base-role-arn=arn:aws:iam::123456789012:role/"
            - "--iptables=true"
            - "--host-ip=$(HOST_IP)"
            - "--node=$(NODE_NAME)"
          env:
            - name: HOST_IP
              valueFrom:
                fieldRef:
                  fieldPath: status.podIP
            - name: NODE_NAME
              valueFrom:
                fieldRef:
                  fieldPath: spec.nodeName
          ports:
            - containerPort: 8181
              hostPort: 8181
              name: http
          securityContext:
            privileged: true

Add an iam.amazonaws.com/role annotation to your pods with the role that you want to assume for this pod. The optional iam.amazonaws.com/external-id will allow the use of an ExternalId as part of the assume role

apiVersion: v1
kind: Pod
metadata:
  name: aws-cli
  labels:
    name: aws-cli
  annotations:
    iam.amazonaws.com/role: role-arn
    iam.amazonaws.com/external-id: external-id
spec:
  containers:
  - image: fstab/aws-cli
    command:
      - "/home/aws/aws/env/bin/aws"
      - "s3"
      - "ls"
      - "some-bucket"
    name: aws-cli

You can use --default-role to set a fallback role to use when annotation is not set.

When creating higher-level abstractions than pods, you need to pass the annotation in the pod template of the resource spec.

Example for a Deployment:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: nginx-deployment
spec:
  replicas: 3
  template:
    metadata:
      annotations:
        iam.amazonaws.com/role: role-arn
      labels:
        app: nginx
    spec:
      containers:
      - name: nginx
        image: nginx:1.9.1
        ports:
        - containerPort: 80

Example for a CronJob:

apiVersion: batch/v1beta1
kind: CronJob
metadata:
  name: my-cronjob
spec:
  schedule: "00 11 * * 2"
  concurrencyPolicy: Forbid
  startingDeadlineSeconds: 3600
  jobTemplate:
    spec:
      template:
        metadata:
          annotations:
            iam.amazonaws.com/role: role-arn
        spec:
          restartPolicy: OnFailure
          containers:
          - name: job
            image: my-image

By using the flag --namespace-restrictions you can enable a mode in which the roles that pods can assume is restricted by an annotation on the pod's namespace. This annotation should be in the form of a json array.

To allow the aws-cli pod specified above to run in the default namespace your namespace would look like the following.

apiVersion: v1
kind: Namespace
metadata:
  annotations:
    iam.amazonaws.com/allowed-roles: |
      ["role-arn"]
  name: default

Note: You can also use glob-based matching for namespace restrictions, which works nicely with the path-based namespacing supported for AWS IAM roles.

Example: to allow all roles prefixed with my-custom-path/ to be assumed by pods in the default namespace, the default namespace would be annotated as follows:

apiVersion: v1
kind: Namespace
metadata:
  annotations:
    iam.amazonaws.com/allowed-roles: |
      ["my-custom-path/*"]
  name: default

If you prefer regexp to glob-based matching you can specify --namespace-restriction-format=regexp, then you can use a regexp in your annotation:

apiVersion: v1
kind: Namespace
metadata:
  annotations:
    iam.amazonaws.com/allowed-roles: |
      ["my-custom-path/.*"]
  name: default

This is the basic RBAC setup to get kube2iam working correctly when your cluster is using rbac. Below is the bare minimum to get kube2iam working.

First we need to make a service account.

---
apiVersion: v1
kind: ServiceAccount
metadata:
  name: kube2iam
  namespace: kube-system

Next we need to setup roles and binding for the the process.

---
apiVersion: v1
items:
  - apiVersion: rbac.authorization.k8s.io/v1
    kind: ClusterRole
    metadata:
      name: kube2iam
    rules:
      - apiGroups: [""]
        resources: ["namespaces","pods"]
        verbs: ["get","watch","list"]
  - apiVersion: rbac.authorization.k8s.io/v1
    kind: ClusterRoleBinding
    metadata:
      name: kube2iam
    subjects:
    - kind: ServiceAccount
      name: kube2iam
      namespace: kube-system
    roleRef:
      kind: ClusterRole
      name: kube2iam
      apiGroup: rbac.authorization.k8s.io
kind: List

You will notice this lives in the kube-system namespace to allow for easier seperation between system services and other services.

Here is what a kube2iam daemonset yaml might look like.

---
apiVersion: apps/v1
kind: DaemonSet
metadata:
  name: kube2iam
  namespace: kube-system
  labels:
    app: kube2iam
spec:
  selector:
    matchLabels:
      name: kube2iam
  template:
    metadata:
      labels:
        name: kube2iam
    spec:
      serviceAccountName: kube2iam
      hostNetwork: true
      containers:
        - image: jtblin/kube2iam:latest
          imagePullPolicy: Always
          name: kube2iam
          args:
            - "--app-port=8181"
            - "--base-role-arn=arn:aws:iam::xxxxxxx:role/"
            - "--iptables=true"
            - "--host-ip=$(HOST_IP)"
            - "--host-interface=weave"
            - "--verbose"
          env:
            - name: HOST_IP
              valueFrom:
                fieldRef:
                  fieldPath: status.podIP
          ports:
            - containerPort: 8181
              hostPort: 8181
              name: http
          securityContext:
            privileged: true

To use kube2iam on OpenShift one needs to configure additional resources.

A complete example for OpenShift 3 looks like this. For OpenShift 4, see the next section.

---
apiVersion: v1
kind: ServiceAccount
metadata:
  name: kube2iam
  namespace: kube-system
---
apiVersion: v1
items:
  - apiVersion: rbac.authorization.k8s.io/v1beta1
    kind: ClusterRole
    metadata:
      name: kube2iam
    rules:
      - apiGroups: [""]
        resources: ["namespaces","pods"]
        verbs: ["get","watch","list"]
  - apiVersion: rbac.authorization.k8s.io/v1beta1
    kind: ClusterRoleBinding
    metadata:
      name: kube2iam
    subjects:
    - kind: ServiceAccount
      name: kube2iam
      namespace: kube-system
    roleRef:
      kind: ClusterRole
      name: kube2iam
      apiGroup: rbac.authorization.k8s.io
kind: List
---
kind: SecurityContextConstraints
apiVersion: v1
metadata:
  name: kube2iam
allowPrivilegedContainer: true
allowHostPorts: true
allowHostNetwork: true
runAsUser:
  type: RunAsAny
seLinuxContext:
  type: MustRunAs
users:
- system:serviceacount:kube-system:kube2iam
---
apiVersion: extensions/v1beta1
kind: DaemonSet
metadata:
  name: kube2iam
  namespace: kube-system
  labels:
    app: kube2iam
spec:
  selector:
    matchLabels:
      name: kube2iam
  template:
    metadata:
      labels:
        name: kube2iam
    spec:
      serviceAccountName: kube2iam
      hostNetwork: true
      nodeSelector:
        role: app
      containers:
        - image: docker.io/jtblin/kube2iam:latest
          imagePullPolicy: Always
          name: kube2iam
          args:
            - "--app-port=8181"
            - "--auto-discover-base-arn"
            - "--iptables=true"
            - "--host-ip=$(HOST_IP)"
            - "--host-interface=tun0"
            - "--verbose"
          env:
            - name: HOST_IP
              valueFrom:
                fieldRef:
                  fieldPath: status.podIP
          ports:
            - containerPort: 8181
              hostPort: 8181
              name: http
          securityContext:
            privileged: true

Note: In (OpenShift) multi-tenancy setups it is recommended to restrict the assumable roles on the namespace level to prevent cross-namespace trust stealing.

To use kube2iam on OpenShift 4, the additional resources are slightly different from those for OpenShift 3 shown above. OpenShift 4 has hard-coded iptables rules that block connections from containers to the EC2 metadata service 169.254.169.254. The kube2iam pods already run with host networking enabled, they are not affected by these OpenShift iptables rules.

The OpenShift iptables rules have implications for pods authenticating through kube2iam though. But let's look at an example for deploying kube2iam on OpenShift 4 first:

---
apiVersion: v1
kind: ServiceAccount
metadata:
  name: kube2iam
  namespace: kube-system
---
apiVersion: v1
items:
  - apiVersion: rbac.authorization.k8s.io/v1beta1
    kind: ClusterRole
    metadata:
      name: kube2iam
    rules:
      - apiGroups: [""]
        resources: ["namespaces","pods"]
        verbs: ["get","watch","list"]
  - apiVersion: rbac.authorization.k8s.io/v1beta1
    kind: ClusterRoleBinding
    metadata:
      name: kube2iam
    subjects:
    - kind: ServiceAccount
      name: kube2iam
      namespace: kube-system
    roleRef:
      kind: ClusterRole
      name: kube2iam
      apiGroup: rbac.authorization.k8s.io
kind: List
---
apiVersion: extensions/v1beta1
kind: DaemonSet
metadata:
  name: kube2iam
  namespace: kube-system
  labels:
    app: kube2iam
spec:
  selector:
    matchLabels:
      name: kube2iam
  template:
    metadata:
      labels:
        name: kube2iam
    spec:
      serviceAccountName: kube2iam
      hostNetwork: true
      nodeSelector:
        node-role.kubernetes.io/worker: ''
      containers:
        - image: docker.io/jtblin/kube2iam:latest
          imagePullPolicy: Always
          name: kube2iam
          args:
            - "--app-port=8181"
            - "--auto-discover-base-arn"
            - "--host-ip=$(HOST_IP)"
            - "--host-interface=tun0"
            - "--verbose"
          env:
            - name: HOST_IP
              valueFrom:
                fieldRef:
                  fieldPath: status.podIP
          ports:
            - containerPort: 8181
              hostPort: 8181
              name: http

Compared to the OpenShift 3 example in the previous section, we removed the kube2iam SecurityContextConstraint. In the kube2iam DaemonSet, we changed the nodeSelector to the match OpenShift 4 worker nodes, removed the iptables argument, and removed the privileged securityContext.

We use the OpenShift hostnetwork SecurityContextConstraint for kube2iam:

oc adm policy add-scc-to-user hostnetwork -n kube-system -z kube2iam

For applications, the iptables rule that kube2iam would create to redirect 169.254.169.254 connections to the kube2iam pods has no effect because the hard-coded iptables rules block those connections on OpenShift 4.

As a workaround, the environment variables http_proxy and no_proxy can be set to use kube2iam as a HTTP proxy when accessing the metadata service. Below is an example for the aws-service-operator:

- kind: Deployment
  apiVersion: apps/v1beta1
  metadata:
    name: aws-service-operator
    namespace: aws-service-operator
  spec:
    replicas: 1
    template:
      metadata:
        annotations:
          iam.amazonaws.com/role: aws-service-operator
        labels:
          app: aws-service-operator
      spec:
        serviceAccountName: aws-service-operator
        containers:
        - name: aws-service-operator
          image: awsserviceoperator/aws-service-operator:v0.0.1-alpha4
          imagePullPolicy: Always
          command:
            - /bin/sh
          args:
          - "-c"
          - export http_proxy=${HOST_IP}:8181; /usr/local/bin/aws-service-operator server --cluster-name=<CLUSTER_NAME> --region=<REGION> --account-id=<ACCOUNT_ID> --k8s-namespace=<K8S_NAMESPACE>
        env:
          - name: HOST_IP
            valueFrom:
              fieldRef:
                apiVersion: v1
                fieldPath: status.hostIP
          - name: no_proxy
            value: "*.amazonaws.com,<KUBE_API_IP>:443"

Compared to the Deployment definition from aws-service-operator/configs/aws-service-operator.yaml, this adds the http_proxy and no_proxy environment variables.

Because we use the IP address of the OpenShift node to access the kube2iam pod, we cannot set http_proxy in the env list, but use a shell command instead.

The value for the no_proxy environment variable is specific to the application. kube2iam only allows proxy connections to 169.254.169.254. All other hostnames or IP addresses that the application connects to through HTTP or HTTPS need to be listed in the no_proxy variable.

For example, the aws-service-operator needs access to various AWS APIs and the Kubernetes API. The Kubernetes API listens on the first IP address in the OpenShift service network. If 172.31.0.0/16 is the OpenShift cluster service network, KUBE_API_IP is 172.31.0.1.

By using the --debug flag you can enable some extra features making debugging easier:

• /debug/store endpoint enabled to dump knowledge of namespaces and role association.

By using the --auto-discover-base-arn flag, kube2iam will auto discover the base ARN via the EC2 metadata service.

By using the --auto-discover-default-role flag, kube2iam will auto discover the base ARN and the IAM role attached to the instance and use it as the fallback role to use when annotation is not set.

STS is a unique service in that it is actually considered a global service that defaults to endpoint at https://sts.amazonaws.com, regardless of your region setting. However, unlike other global services (e.g. CloudFront, IAM), STS also has regional endpoints which can only be explicitly used programatically. The use of a regional sts endpoint can reduce the latency for STS requests.

kube2iam supports the use of STS regional endpoints by using the --use-regional-sts-endpoint flag as well as by setting the appropriate AWS_REGION environment variable in your daemonset environment. With these two settings configured, kube2iam will use the STS api endpoint for that region. If you enable debug level logging, the sts endpoint used to retrieve credentials will be logged.

kube2iam exports a number of Prometheus metrics to assist with monitoring the system's performance. By default, these are exported at the /metrics HTTP endpoint on the application server port (specified by --app-port). This does not always make sense, as anything with access to the application server port can assume roles via kube2iam. To mitigate this use the --metrics-port argument to specify a different port that will host the /metrics endpoint.

All of the exported metrics are prefixed with kube2iam_. See the Prometheus documentation for more information on how to get up and running with Prometheus.

By default, kube2iam will use the in-cluster method to connect to the kubernetes master, and use the iam.amazonaws.com/role annotation to retrieve the role for the container. Either set the base-role-arn option to apply to all roles and only pass the role name in the iam.amazonaws.com/role annotation, otherwise pass the full role ARN in the annotation.

$ kube2iam --help
Usage of kube2iam:
      --api-server string                     Endpoint for the api server
      --api-token string                      Token to authenticate with the api server
      --app-port string                       Kube2iam server http port (default "8181")
      --auto-discover-base-arn                Queries EC2 Metadata to determine the base ARN
      --auto-discover-default-role            Queries EC2 Metadata to determine the default Iam Role and base ARN, cannot be used with --default-role, overwrites any previous setting for --base-role-arn
      --backoff-max-elapsed-time duration     Max elapsed time for backoff when querying for role. (default 2s)
      --backoff-max-interval duration         Max interval for backoff when querying for role. (default 1s)
      --base-role-arn string                  Base role ARN
      --iam-role-session-ttl                  Length of session when assuming the roles (default 15m)
      --debug                                 Enable debug features
      --default-role string                   Fallback role to use when annotation is not set
      --host-interface string                 Host interface for proxying AWS metadata (default "docker0")
      --host-ip string                        IP address of host
      --iam-role-key string                   Pod annotation key used to retrieve the IAM role (default "iam.amazonaws.com/role")
      --iam-external-id string                Pod annotation key used to retrieve the IAM ExternalId (default "iam.amazonaws.com/external-id")
      --insecure                              Kubernetes server should be accessed without verifying the TLS. Testing only
      --iptables                              Add iptables rule (also requires --host-ip)
      --log-format string                     Log format (text/json) (default "text")
      --log-level string                      Log level (default "info")
      --metadata-addr string                  Address for the ec2 metadata (default "169.254.169.254")
      --metrics-port string                   Metrics server http port (default: same as kube2iam server port) (default "8181")
      --namespace-key string                  Namespace annotation key used to retrieve the IAM roles allowed (value in annotation should be json array) (default "iam.amazonaws.com/allowed-roles")
      --cache-resync-period                   Refresh interval for pod and namespace caches
      --resolve-duplicate-cache-ips           Queries the k8s api server to find the source of truth when the pod cache contains multiple pods with the same IP
      --namespace-restriction-format string   Namespace Restriction Format (glob/regexp) (default "glob")
      --namespace-restrictions                Enable namespace restrictions
      --node string                           Name of the node where kube2iam is running
      --use-regional-sts-endpoint             use the regional sts endpoint if AWS_REGION is set
      --verbose                               Verbose
      --version                               Print the version and exits

• Use minikube to run cluster locally
• Build and push dev image to docker hub: make docker-dev DOCKER_REPO=<your docker hub username>
• Update deployment.yaml as needed
• Deploy to local kubernetes cluster: kubectl create -f deployment.yaml or kubectl delete -f deployment.yaml && kubectl create -f deployment.yaml
• Expose as service: kubectl expose deployment kube2iam --type=NodePort
• Retrieve the services url: minikube service kube2iam --url
• Test your changes e.g. curl -is $(minikube service kube2iam --url)/healthz

Jerome Touffe-Blin, @jtblin, About me

kube2iam is copyright 2020 Jerome Touffe-Blin and contributors. It is licensed under the BSD license. See the included LICENSE file for details.