fix(megaraid): rescan SCSI bus after creating a virtual drive

[ezekiel-alexrod]

Problem

Creating several MegaRAID virtual drives back to back could fail with:

failed to find new logical volume: failed to get logical volumes: failed to get logical volume <id>: failed to get paths: failed to compute paths from physical drive: failed to compute permanent path

createLV resolves the freshly created volume through its /dev/disk/by-id/wwn-0x<naa> symlink. When many volumes are created in a row, that symlink was sometimes missing at resolution time and appeared a few seconds later, so the call failed non-deterministically.

Root cause

Not a udev latency issue. Instrumenting the sequence showed:

• storcli add vd returns before the OS has a block device for the volume.
• The megaraid_sas driver auto-announces new VDs via async events, but only for the first few created in rapid succession. After that, VDs exist on the controller with no /dev/sdX at all.
• Once a device is finally discovered, its by-id symlink appears in the same instant, so udevadm settle is a no-op here (it returned in single-digit ms every time).

Measured per-volume discovery latency: ~1s for the first volumes (auto discovery), then no discovery until an explicit SCSI bus rescan, after which the device and its symlink appear together.

Fix

After add vd, force a targeted SCSI bus rescan and poll findNewLogicalVolume until the new volume resolves (bounded by a 60s timeout, 1s interval):

• Rescan writes - - - only to hosts whose proc_name is megaraid_sas. Scanning every host walks empty SATA ports and adds multi-second link down timeouts, so it is deliberately scoped to the MegaRAID hosts.
• The rescan is best effort: its error is only surfaced if the poll times out.
• CustomRescanSCSIHosts is a package-level seam (like CustomFileExists) so the sysfs side effect can be stubbed in tests.

Testing

• New unit test TestCreateLVForcesRescanAndSettles: reproduces the failure on the first resolution attempt (matches the exact production error chain via the single-drive ComputePaths fallback), verifies createLV retries after a rescan and returns the volume, and asserts a rescan was triggered.
• go build ./..., full go test ./..., go vet ./... and golangci-lint all clean.
• Verified end to end on hardware: a multi-volume creation run that previously failed now completes.

[claude]

• megaraid_test.go:720 — TestCreateLVForcesRescanAndSettles sleeps 1s unnecessarily; the settle timeout/interval vars were made package-level for test shrinking but the test doesn't use that seam.
  
  Everything else looks solid: error handling follows the Wrap/Wrapf convention, the rescan is properly scoped to megaraid_sas hosts, the CustomRescanSCSIHosts seam mirrors the existing CustomFileExists pattern, and the test correctly exercises the retry path through the FileExists once-false mock.
  
  Review by Claude Code

[claude]

• logicalvolume.go:41-42 — newVolumeSettleTimeout and newVolumeSettleInterval are unexported but the tests use package megaraid_test (external), so they cannot be shrunk as the comment claims. Export them to match the CustomRescanSCSIHosts / CustomFileExists pattern, and shrink them in TestCreateLVForcesRescanAndSettles to avoid a 1 s sleep per run.
  
  Overall: solid fix for a real hardware-level race. The retry loop, scoped rescan, and error surfacing are well designed. The only issue is the test timing oversight above.
  
  Review by Claude Code

[claude]

LGTM — clean fix for a real production race. The retry loop, targeted SCSI rescan, and test seam all follow the existing patterns (CustomFileExists / CustomEvalSymlinks). Error handling properly surfaces rescan failures on timeout via lastRescanErr.

Review by Claude Code