Skip to content

Proxmox CSI driver

Optional, for every install method: PersistentVolumes become Proxmox disks that are hot-plugged into the VM of the node the pod runs on, with the Proxmox CSI plugin.

# group_vars/all/install.yml
proxmox_csi_enabled: true
proxmox_csi_storage: local-zfs          # any Proxmox storage with content "images"
proxmox_csi_storage_class: proxmox-data
proxmox_csi_default_class: true
proxmox_csi_fstype: xfs

Installed at the end of ./ocp.sh deploy, or on a running cluster:

./ocp.sh addons                          # = ansible-playbook addons.yml

What gets installed

namespace csi-proxmox privileged pod security; controller + node SAs get the privileged SCC
Helm release proxmox-csi-plugin controller on the masters, node plugin on every node (masters too)
StorageClass proxmox-data WaitForFirstConsumer, Delete, xfs, ssd=1
CronJob proxmox-topology-labels not on external: sets region/zone labels per node (every 2 min), incl. autoscaled workers

How a volume finds its VM

The driver needs to know which VM a node is and on which Proxmox node it runs:

  • external: from the CCM's providerID and zone label.
  • agent / ipi (platform baremetal): the VM whose name starts with the node name and whose SMBIOS UUID matches the node's system UUID. The zone label comes from the CronJob, from the same node → Proxmox node map Shiftmox uses to create the VMs.

Good to know

  • Local storage pins volumes. On local-zfs / lvm-thin a volume lives on one Proxmox node; its pod can only run on VMs on that node. Shared storage (Ceph RBD, NFS) has no such limit.
  • The VMs already use the required VirtIO SCSI single controller.
  • Token: the in-cluster token (proxmox_k8s_token_*, default the autoscale token) needs VM.Audit, VM.Config.Disk, Datastore.Allocate, Datastore.AllocateSpace, Datastore.Audit — the deploy token has them.
  • Disks show up in the VM's hardware list as scsiN; don't remove them there while a pod uses them.