feat: add lblk (non-NVMe block device) support for storage nodes

[boddumanohar]

fixes: #436

Exposes sbcli's lblk device mode as a cluster-level choice plus per-fleet device selectors, so storage nodes can run on generic Linux block devices — cloud block volumes, virtio-blk disks, SAS/SATA disks — instead of requiring NVMe PCIe hardware. In lblk mode each selected device is wrapped in an SPDK AIO bdev; everything from alceml upward is unchanged.

apiVersion: storage.simplyblock.io/v1alpha2
kind: StorageCluster
spec:
  deviceClass: LogicalBlock       # cluster-wide, immutable, defaults to NVMe
---
apiVersion: storage.simplyblock.io/v1alpha1
kind: StorageNodeSet
spec:
  enableLblk: true
  blkSerials: ["QEMU_HARDDISK_drive-scsi1", "QEMU_HARDDISK_drive-scsi2"]
  blkForceFormat: false          # destructive; reuses a partitioned disk
  lblkJournalPercent: 3          # only when the smallest selection is a partition

• deviceClass (NVMe|LogicalBlock, default NVMe) is the cluster-wide choice, mapped onto sbcli's device_mode (nvme|lblk) on POST /api/v2/clusters/. Immutable — the backend bakes it in at cluster-create and cannot switch a live cluster. This field predates this PR: it shipped on StorageCluster v1alpha2 in feat(operator): StorageCluster moves to v1alpha2, and its operations are rebuilt #536, and this PR wires it through to cluster-create rather than adding a field of its own.
• enableLblk turns a fleet onto block devices. Immutable, and mutually exclusive with the NVMe selectors.
• blkNames / blkNamesExclude / blkSerials select the devices: by kernel name, by kernel name negatively, or by serial/WWN. At most one may be set; with none set, every eligible whole disk on the node is used. blkSerials is the one to prefer on cloud and virtualized disks, whose /dev/sdX name is not stable across reboots.
• blkForceFormat wipes partition tables and filesystem signatures so a partitioned disk becomes eligible as a whole disk. Destructive, and off by default — without it a partitioned disk is simply rejected as ineligible.
• lblkJournalPercent carves the journal out of the smallest selected device when that device is a partition rather than a whole disk. Ignored otherwise; the backend's own default is 3.

Every field exists on StorageNodeSet.spec, on its per-node nodeConfigs entries, and on StorageNode.spec.overrides, exactly like the existing PCIe selectors, and the most specific one wins.

How a selection reaches the node

Device selection is not part of the "add storage node" REST API — it happens node-locally in node_configure.py, which the DaemonSet init container already drives for NVMe. The lblk fields ride the same path rather than introducing a second integration point:

Field	ConfigMap env	node_configure.py flag
enableLblk	LBLK	--lblk
blkNames	BLK_NAMES	--blk-names
blkNamesExclude	BLK_NAMES_EXCLUDE	--blk-names-exclude
blkSerials	BLK_SERIALS	--blk-serials
blkForceFormat	LBLK_FORCE_FORMAT	--force-format
lblkJournalPercent	LBLK_JM_PERCENT	--jm-percent

The s-node-api-config-generator init container that runs node_configure.py now also mounts the host's /dev and /sys (already mounted on the main container, for SPDK). NVMe discovery never needed them: it resolves a device to a PCI address purely from /sys/class/nvme and /sys/bus/pci, sysfs metadata that's visible in any container without a hostPath mount. lblk devices have no such bus to walk — eligibility (mounted, held by LVM/md, partitioned, NUMA node) can only be read off the real block layer: lsblk over the actual device nodes, /sys/block/<name>/holders, and /dev/disk/by-id/* for stable identity. That by-id tree in particular only exists because the host's udev populates it; a container's own default /dev never has one, hostPath or not.

What admission rejects

Change	Result
lblk selector and PCIe selector on one StorageNodeSet	❌ rejected — the backend rejects a node configured with both
blkNames/blkNamesExclude/blkSerials set on a new object	✅ allowed
the same fields edited after they are set	❌ rejected as immutable
deviceClass changed on a live cluster	❌ rejected as immutable (generated from StorageCluster's +k8s:immutable marker)

The immutability covers the unset-to-set transition too, which matters more than it looks: a fleet commonly starts with no selector at all (auto-select), and node_configure.py silently skips regenerating config for an already-provisioned node — so an edit that Kubernetes accepted would read back as though it had taken effect while changing nothing.

Scope

This PR is now field plumbing plus the two fixes specific to lblk itself (admission's immutability/mutual-exclusivity rules and the config-generator's /dev//sys mounts). Two other fixes found during the same config-israel testing pass but unrelated to lblk — a StorageNode deletion race when its parent StorageNodeSet is already gone, and the storage node ServiceAccount token's expiration — are split out to #535 rather than riding along here.

Rebase note (2026-09-15)

Rebased onto main past the StorageCluster v1alpha2 rebuild (#536). That
rebuild had already shipped spec.deviceClass: NVMe|LogicalBlock — the same
concept this PR originally added as its own spec.deviceMode on v1alpha1 —
and v1alpha1 is now a spoke converting into the v1alpha2 hub, so a
deviceMode field added only there would round-trip through the conversion
and be silently dropped. This PR now drops that field and instead wires the
existing deviceClass through to cluster-create in the new
storagecluster_controller.go. StorageNodeSet/StorageNode's lblk fields
are unaffected: those kinds have not moved to v1alpha2 yet.

🤖 Generated with Claude Code