Provisioning Identity through SDS

 7 minute read

This task shows how to enable SDS (secret discovery service) for Istio identity provisioning.

Prior to Istio 1.1, the keys and certificates of Istio workloads were generated by Citadel and distributed to sidecars through secret-volume mounted files, this approach has the following minor drawbacks:

  • Performance regression during certificate rotation: When certificate rotation happens, Envoy is hot restarted to pick up the new key and certificate, causing performance regression.

  • Potential security vulnerability: The workload private keys are distributed through Kubernetes secrets, with known risks.

These issues are addressed in Istio 1.1 through the SDS identity provision flow. The workflow can be described as follows.

  1. The workload sidecar Envoy requests the key and certificates from the Citadel agent: The Citadel agent is a SDS server, which runs as per-node DaemonSet. In the request, Envoy passes a Kubernetes service account JWT to the agent.

  2. The Citadel agent generates a key pair and sends the CSR request to Citadel: Citadel verifies the JWT and issues the certificate to the Citadel agent.

  3. The Citadel agent sends the key and certificate back to the workload sidecar.

This approach has the following benefits:

  • The private key never leaves the node: It is only in the Citadel agent and Envoy sidecar’s memory.

  • The secret volume mount is no longer needed: The reliance on the Kubernetes secrets is eliminated.

  • The sidecar Envoy is able to dynamically renew the key and certificate through the SDS API: Certificate rotations no longer require Envoy to restart.

Before you begin

  • Set up Istio by following the instructions using Helm with SDS setup and global mutual TLS enabled:

    Zip
    $ kubectl create namespace istio-system
$ for i in install/kubernetes/helm/istio-init/files/crd*yaml; do kubectl apply -f $i; done
$ helm template install/kubernetes/helm/istio --name istio --namespace istio-system --values @install/kubernetes/helm/istio/values-istio-sds-auth.yaml@ > istio-auth-sds.yaml
$ kubectl apply -f istio-auth-sds.yaml

Service-to-service mutual TLS using key/certificate provisioned through SDS

Follow the authentication policy task to setup test services.

ZipZipZipZip
$ kubectl create ns foo
$ kubectl apply -f <(istioctl kube-inject -f @samples/httpbin/httpbin.yaml@) -n foo
$ kubectl apply -f <(istioctl kube-inject -f @samples/sleep/sleep.yaml@) -n foo
$ kubectl create ns bar
$ kubectl apply -f <(istioctl kube-inject -f @samples/httpbin/httpbin.yaml@) -n bar
$ kubectl apply -f <(istioctl kube-inject -f @samples/sleep/sleep.yaml@) -n bar

Verify all mutual TLS requests succeed:

$ for from in "foo" "bar"; do for to in "foo" "bar"; do kubectl exec $(kubectl get pod -l app=sleep -n ${from} -o jsonpath={.items..metadata.name}) -c sleep -n ${from} -- curl "http://httpbin.${to}:8000/ip" -s -o /dev/null -w "sleep.${from} to httpbin.${to}: %{http_code}\n"; done; done
sleep.foo to httpbin.foo: 200
sleep.foo to httpbin.bar: 200
sleep.bar to httpbin.foo: 200
sleep.bar to httpbin.bar: 200

Verifying no secret-volume mounted file is generated

To verify that no secret-volume mounted file is generated, access the deployed workload sidecar container:

$ kubectl exec -it $(kubectl get pod -l app=sleep -n foo -o jsonpath={.items..metadata.name}) -c istio-proxy -n foo  -- /bin/bash

As you can see there is no secret file mounted at /etc/certs folder.

Increasing security with pod security policies

The Istio Secret Discovery Service (SDS) uses the Citadel agent to distribute the certificate to the Envoy sidecar via a Unix domain socket. All pods running in the same Kubernetes node share the Citadel agent and Unix domain socket.

To prevent malicious modifications to the Unix domain socket, enable the pod security policy to restrict the pod’s permission on the Unix domain socket. Otherwise, a malicious pod could hijack the Unix domain socket to break the SDS service or steal the identity credentials from other pods running on the same Kubernetes node.

To enable the pod security policy, perform the following steps:

  1. The Citadel agent fails to start unless it can create the required Unix domain socket. Apply the following pod security policy to only allow the Citadel agent to modify the Unix domain socket:

    $ cat <<EOF | kubectl apply -f -
apiVersion: extensions/v1beta1
kind: PodSecurityPolicy
metadata:
  name: istio-nodeagent
spec:
  allowedHostPaths:
  - pathPrefix: "/var/run/sds"
  seLinux:
    rule: RunAsAny
  supplementalGroups:
    rule: RunAsAny
  runAsUser:
    rule: RunAsAny
  fsGroup:
    rule: RunAsAny
  volumes:
  - '*'
---
kind: Role
apiVersion: rbac.authorization.k8s.io/v1
metadata:
  name: istio-nodeagent
  namespace: istio-system
rules:
- apiGroups:
  - extensions
  resources:
  - podsecuritypolicies
  resourceNames:
  - istio-nodeagent
  verbs:
  - use
---
apiVersion: rbac.authorization.k8s.io/v1
kind: RoleBinding
metadata:
  name: istio-nodeagent
  namespace: istio-system
roleRef:
  apiGroup: rbac.authorization.k8s.io
  kind: Role
  name: istio-nodeagent
subjects:
- kind: ServiceAccount
  name: istio-nodeagent-service-account
  namespace: istio-system
EOF

  2. To stop other pods from modifying the Unix domain socket, change the allowedHostPaths configuration for the the path the Citadel agent uses for the Unix domain socket to readOnly: true.

    $ cat <<EOF | kubectl apply -f -
apiVersion: extensions/v1beta1
kind: PodSecurityPolicy
metadata:
  name: istio-sds-uds
spec:
 # Protect the unix domain socket from unauthorized modification
 allowedHostPaths:
 - pathPrefix: "/var/run/sds"
   readOnly: true
 # Allow the istio sidecar injector to work
 allowedCapabilities:
 - NET_ADMIN
 seLinux:
   rule: RunAsAny
 supplementalGroups:
   rule: RunAsAny
 runAsUser:
   rule: RunAsAny
 fsGroup:
   rule: RunAsAny
 volumes:
 - '*'
---
kind: ClusterRole
apiVersion: rbac.authorization.k8s.io/v1
metadata:
  name: istio-sds-uds
rules:
- apiGroups:
  - extensions
  resources:
  - podsecuritypolicies
  resourceNames:
  - istio-sds-uds
  verbs:
  - use
---
apiVersion: rbac.authorization.k8s.io/v1
kind: ClusterRoleBinding
metadata:
  name: istio-sds-uds
roleRef:
  apiGroup: rbac.authorization.k8s.io
  kind: ClusterRole
  name: istio-sds-uds
subjects:
- apiGroup: rbac.authorization.k8s.io
  kind: Group
  name: system:serviceaccounts
EOF

  3. Enable pod security policies for your platform. Each supported platform enables pod security policies differently. Please refer to the pertinent documentation for your platform. If you are using the Google Kubernetes Engine (GKE), you must enable the pod security policy controller.

  4. Run the following command to restart the Citadel agents:

    $ kubectl delete pod -l 'app=nodeagent' -n istio-system
pod "istio-nodeagent-dplx2" deleted
pod "istio-nodeagent-jrbmx" deleted
pod "istio-nodeagent-rz878" deleted

  5. To verify that the Citadel agents work with the enabled pod security policy, wait a few seconds and run the following command to confirm the agents started successfully:

    $ kubectl get pod -l 'app=nodeagent' -n istio-system
NAME                    READY   STATUS    RESTARTS   AGE
istio-nodeagent-p4p7g   1/1     Running   0          4s
istio-nodeagent-qdwj6   1/1     Running   0          5s
istio-nodeagent-zsk2b   1/1     Running   0          14s

  6. Run the following command to start a normal pod.

    $ cat <<EOF | kubectl apply -f -
apiVersion: extensions/v1beta1
kind: Deployment
metadata:
  name: normal
spec:
  replicas: 1
  template:
    metadata:
      labels:
        app: normal
    spec:
      containers:
      - name: normal
        image: pstauffer/curl
        command: ["/bin/sleep", "3650d"]
        imagePullPolicy: IfNotPresent
EOF

  7. To verify that the normal pod works with the pod security policy enabled, wait a few seconds and run the following command to confirm the normal pod started successfully.

    $ kubectl get pod -l 'app=normal'
NAME                      READY   STATUS    RESTARTS   AGE
normal-64c6956774-ptpfh   2/2     Running   0          8s

  8. Start a malicious pod that tries to mount the Unix domain socket using a write permission.

    $ cat <<EOF | kubectl apply -f -
apiVersion: extensions/v1beta1
kind: Deployment
metadata:
  name: malicious
spec:
  replicas: 1
  template:
    metadata:
      labels:
        app: malicious
    spec:
      containers:
      - name: malicious
        image: pstauffer/curl
        command: ["/bin/sleep", "3650d"]
        imagePullPolicy: IfNotPresent
        volumeMounts:
        - name: sds-uds
          mountPath: /var/run/sds
      volumes:
      - name: sds-uds
        hostPath:
          path: /var/run/sds
          type: ""
EOF

  9. To verify that the Unix domain socket is protected, run the following command to confirm the malicious pod failed to start due to the pod security policy:

    $ kubectl describe rs -l 'app=malicious' | grep Failed
Pods Status:    0 Running / 0 Waiting / 0 Succeeded / 0 Failed
  ReplicaFailure   True    FailedCreate
  Warning  FailedCreate  4s (x13 over 24s)  replicaset-controller  Error creating: pods "malicious-7dcfb8d648-" is forbidden: unable to validate against any pod security policy: [spec.containers[0].volumeMounts[0].readOnly: Invalid value: false: must be read-only]

Cleanup

  1. Clean up the test services and the Istio control plane:

    $ kubectl delete ns foo
$ kubectl delete ns bar
$ kubectl delete -f istio-auth-sds.yaml

  2. Disable the pod security policy in the cluster using the documentation of your platform. If you are using GKE, disable the pod security policy controller.

  3. Delete the pod security policy and the test deployments:

    $ kubectl delete psp istio-sds-uds istio-nodeagent
$ kubectl delete role istio-nodeagent -n istio-system
$ kubectl delete rolebinding istio-nodeagent -n istio-system
$ kubectl delete clusterrole istio-sds-uds
$ kubectl delete clusterrolebinding istio-sds-uds
$ kubectl delete deploy malicious
$ kubectl delete deploy normal

Caveats

Currently, the SDS identity provision flow has the following caveats:

  • You still need secret volume mount for enabling the control plane security. Enabling SDS for the control plane security remains a work in progress.

  • Smoothly migrating a cluster from using secret volume mount to using SDS is a work in progress.