Menu
On this page

Real-kernel qualification#

The M0 and M1a device stages cover a null target, a RAM target, and 1/2/4/8 queues. The corrected implementation passed the device stages on the kernels below. The validation record identifies the earlier failed revisions and the scope of each reported result.

Device-stage results recorded 2026-10-10. Validation evidence and scope.

Kernel familyKernelM0/M1a device stages
Ubuntu6.8pass
Debian6.12pass
Ubuntu7.0pass
Mainline7.2.6pass
Fedora7.2.9pass
Mainline7.3-rc3pass

The full core runbook passed 33/33 stages on Ubuntu 6.8 and Debian 6.12. Every binding passed its RAM data-integrity stages on Linux 6.12. Binding coverage on the other kernels hasn’t been established by these results.

The runbook checks direct write/read comparisons, CRC32C fio verification, lifecycle cleanup, leaked devices and mappings, and new kernel warnings. A skip or timeout doesn’t count as a pass. RAM qualification covers volatile storage; it doesn’t establish durability across a crash.

Seeded property campaign#

The committed campaign ledger records independent seeds for each case, with these results:

PropertyCasesFailures
Request decoding200,0000
Completion sequences200,0000
Lifecycle against a model200,0000
Versioned record byte round trips200,0000

These are bounded randomized properties with replayable seeds. They supplement deterministic models and real-kernel testing. Coverage-guided fuzzing isn’t qualified by this campaign.

Defects exposed by real kernels#

The validation record records these fixes:

DefectCorrection
Failed START left primed FETCH commands to drainRun STOP after partial priming or failed START, and keep the owners driving until abort completions drain.
FLUSH’s sentinel sector overflowed address arithmeticDecode FLUSH without treating its wire range as an addressed payload.
Independent character-device opens failed with multiple queuesOpen once, and duplicate the surviving file description for later queues.
Additive capability flags returned by newer kernels were rejectedAdmit compatible capability flags while continuing to reject unrequested I/O modes and unknown bits.
A udev probe raced a held-request testArm the hold after client open, and match the intended direct request exactly.

The comparative benchmarks have a separate measurement scope. The kernel bug guide tracks findings in the kernel itself.