Kubernetes supports the 2 volume modes for PVC: Filesystem and Block. When a pod defines the security context and requests a Longhorn PVC, Kubernetes will handle the ownership and permission modification for the PVC differently based on the volume mode.

Longhorn PVC with Filesystem Volume Mode

Because the Longhorn CSI driver csiDriver.spec.fsGroupPolicy is set to ReadWriteOnceWithFSType, the Kubelet attempts to change the ownership and permission of a Longhorn PVC in the following manner:

1. Check pod.spec.securityContext.fsGroup.
   • If non-empty, continue to the next step.
   • If empty, the Kubelet doesn’t attempt to change the ownership and permission for the volume.
2. Check fsType of the PV and accessModes of the PVC.
   • If the PV’s fsType is defined and the PVC’s accessModes list contains ReadWriteOnly, continue to the next step.
   • Otherwise, the Kubelet doesn’t attempt to change the ownership and permission for the volume.
3. Check pod.spec.securityContext.fsGroupChangePolicy.
   • If the pod.spec.securityContext.fsGroupChangePolicy is set to always or empty, the kubelet performs the following actions:
     • Ensures that all processes of the containers inside the pod are part of the supplementary group id pod.spec.securityContext.fsGroup
     • Ensures that any new files created in the volume will be in group id pod.spec.securityContext.fsGroup
     • Recursively changes permission and ownership of the volume to have the same group id as pod.spec.securityContext.fsGroup every time the volume is mounted
   • If the pod.spec.securityContext.fsGroupChangePolicy is set to OnRootMismatch:
     • If the root of the volume already has the correct permissions (i.e., belongs to the group id as pod.spec.securityContext.fsGroup) , the recursive permission and ownership change will be skipped.
     • Otherwise, Kubelet recursively changes permission and ownership of the volume to have the same group id as pod.spec.securityContext.fsGroup

For more information, see:

• https://kubernetes.io/docs/tasks/configure-pod-container/security-context/#configure-volume-permission-and-ownership-change-policy-for-pods
• https://github.com/longhorn/longhorn/issues/2131#issuecomment-778897129

Longhorn PVC with Block Volume Mode

For PVC with Block volume mode, Kubelet never attempts to change the permission and ownership of the block device when making it available inside the container. You must set the correct group ID in the pod.spec.securityContext for the pod to be able to read and write to the block device or run the container as root.

By default, Longhorn puts the block device into group id 6, which is typically associated with the “disk” group. Therefore, pods that use Longhorn PVC with Block volume mode must either set the group id 6 in the pod.spec.securityContext, or run as root. For example:

1. Pod that sets the group id 6 in the pod.spec.securityContext

   apiVersion: v1
   kind: PersistentVolumeClaim
   metadata:
     name: longhorn-block-vol
   spec:
     accessModes:
       - ReadWriteOnce
     volumeMode: Block
     storageClassName: longhorn
     resources:
       requests:
         storage: 2Gi
   ---
   apiVersion: v1
   kind: Pod
   metadata:
     name: block-volume-test
     namespace: default
   spec:
     securityContext:
       runAsGroup: 1000
       runAsNonRoot: true
       runAsUser: 1000
       supplementalGroups:
       - 6
     containers:
       - name: block-volume-test
         image: ubuntu:20.04
         command: ["sleep", "360000"]
         imagePullPolicy: IfNotPresent
         volumeDevices:
           - devicePath: /dev/longhorn/testblk
             name: block-vol
     volumes:
       - name: block-vol
         persistentVolumeClaim:
           claimName: longhorn-block-vol
   

2. Pod that runs as root

   apiVersion: v1
   kind: PersistentVolumeClaim
   metadata:
     name: longhorn-block-vol
   spec:
     accessModes:
       - ReadWriteOnce
     volumeMode: Block
     storageClassName: longhorn
     resources:
       requests:
         storage: 2Gi
   ---
   apiVersion: v1
   kind: Pod
   metadata:
     name: block-volume-test
     namespace: default
   spec:
     containers:
       - name: block-volume-test
         image: ubuntu:20.04
         command: ["sleep", "360000"]
         imagePullPolicy: IfNotPresent
         volumeDevices:
           - devicePath: /dev/longhorn/testblk
             name: block-vol
     volumes:
       - name: block-vol
         persistentVolumeClaim:
           claimName: longhorn-block-vol