iSCSILogicalUnit: add block size override and backstore type selection

[ryan-ronnander]

Summary

This PR adds block size override and backstore type selection to the iSCSILogicalUnit resource agent for the lio-t (targetcli) and scst implementations.

Commits

• block_size parameter — Override the block size (512, 1024, 2048, or 4096) presented to iSCSI initiators. No-op when unset; the kernel default is preserved.

  • lio-t: writes to the LIO configfs block_size attribute after backstore creation, before LUN mapping
  • scst: passes blocksize in the scstadmin -open_dev attributes string at device creation time

• scst_bstype parameter — Select the SCST device handler (vdisk_blockio or vdisk_fileio). The handler was previously hardcoded to vdisk_blockio. The vdisk_fileio handler routes I/O through the kernel page cache, which is required when presenting smaller block sizes on 4K-native backing devices (for example, block_size=512 on a 4K NVMe drive).

• Unsupported parameter warnings — Wire scst_bstype into the unsupported parameter lists for non-scst implementations, and liot_bstype into the scst list, so users get a warning when setting parameters that do not apply to their implementation.

• Fix lio-t configfs paths — The lio-t code path hardcoded iblock_* in configfs paths for scsi_sn, product_id, block_size, and emulate_* attribute writes. This only works for the default block backstore. When liot_bstype is fileio or pscsi, the configfs directory is fileio_* or pscsi_* respectively. Both start and monitor functions now resolve the correct path based on liot_bstype.

Allows these new parameters to be used in:

• Pacemaker
• DRBD Reactor - Reactor uses available resource agents from this repo.
• LINSTOR Gateway - A development build successfully passes these parameters through to drbd-reactor promoter configs, which invoke the resource agent.

Testing

Tested on AlmaLinux 9, kernel 5.14.0-611.45.1.el9_7.x86_64, SCST master (5d74814).

• lio-t block backstore with block_size=4096: start, monitor, stop, Pacemaker failover
• lio-t fileio backstore with block_size=4096: start, monitor, stop via LINSTOR Gateway
• scst vdisk_blockio with block_size=4096: full target and LU lifecycle
• scst vdisk_fileio with block_size=512: full target and LU lifecycle
• All four block sizes (512, 1024, 2048, 4096) verified on both lio-t and scst
• Initiator verification via blockdev --getss and blockdev --getpbsz
• Data integrity (64MB sha256 round-trip) on lio-t block backstore with block_size=4096

[knet-jenkins]

Can one of the project admins check and authorise this run please: https://haci.fast.eng.rdu2.dc.redhat.com/job/resource-agents/job/resource-agents-pipeline/job/PR-2148/1/input

[knet-jenkins]

Can one of the project admins check and authorise this run please: https://haci.fast.eng.rdu2.dc.redhat.com/job/resource-agents/job/resource-agents-pipeline/job/PR-2148/2/input

[knet-jenkins]

Can one of the project admins check and authorise this run please: https://haci.fast.eng.rdu2.dc.redhat.com/job/resource-agents/job/resource-agents-pipeline/job/PR-2148/3/input

[ryan-ronnander]

While the original commit covers the typical use case of presenting 4K block sizes to initiators on 512-native storage (512 → 4K), doing the opposite (4K → 512) requires a fileio backstore. The liot_bstype parameter already exposed this for targetcli, but SCST had no equivalent (the handler was hardcoded to vdisk_blockio).

The second commit adds scst_bstype to allow selecting vdisk_fileio, and the third wires it into the unsupported parameter warnings for consistency.

Tested both directions on AlmaLinux 9 with targetcli and SCST. Happy with where this MR is.

EDIT: A few more tweaks were needed to targetcli fileio backend handling due to hardcoded configfs paths - 146da1e.

[knet-jenkins]

Can one of the project admins check and authorise this run please: https://haci.fast.eng.rdu2.dc.redhat.com/job/resource-agents/job/resource-agents-pipeline/job/PR-2148/4/input

[knet-jenkins]

Can one of the project admins check and authorise this run please: https://haci.fast.eng.rdu2.dc.redhat.com/job/resource-agents/job/resource-agents-pipeline/job/PR-2148/5/input

[ryan-ronnander]

Needed to expose yet another parameter in 8f9f820 to ensure we can set the correct write-back/write-through caching when used in a DRBD context.

• Added emulate_write_cache in the latest commit. The fileio backstore defaults to write-back caching (emulate_write_cache=1), which is unsafe for HA configurations where data must be flushed to disk before the iSCSI write is acknowledged. This follows the existing emulate_tpu, emulate_3pc, emulate_caw pattern.
• LINSTOR Gateway uses these parameters to manage HA iSCSI targets over DRBD-replicated storage. When Gateway selects the fileio backstore (unreleased dev build, needed for presenting smaller block sizes on 4K-native backing devices), it automatically sets emulate_write_cache=0 to ensure write-through mode.
• SCST is unaffected — its vdisk_fileio handler already uses write_through=1 at device creation time, which is hardcoded in the RA.

[oalbrigt]

Nice. I'll review it later this week or next week.