← work systems · networking · observed

197 passing tests were not evidence that RDMA worked.

roce-preflight is a diagnostic tool for RoCE hosts. Its most useful result was finding the boundary its own test suite had not crossed.

Diagram showing 197 passing unit tests, a real Soft-RoCE verbs path, and four defects exposed by device execution.
Observed evidence. The important transition is not 197 → 208 tests; it is synthetic state → independent Linux RDMA state.
Problem

Parsers, ranking logic, snapshots, and lifecycle decisions were well tested against synthetic state, but the suite did not require an RDMA device.

Result

Real RDMA execution exposed four defects after 197 tests passed. The first attempted correction also passed the unit suite and still failed on-device.

Independent evidence

The CI job creates a Linux soft-RoCE device, asserts the backend did not fall back to simulation, exercises real verbs/sysfs state, and sends real perftest traffic.

Boundary

Soft-RoCE validates kernel/userspace interface assumptions. It does not establish mlx5, firmware, PCIe/NUMA, switch, PFC/ECN, multi-host, or production-performance claims.

What failed

MTUlink rate was printed where active MTU belonged
GID diagnosisthe report could contradict the live GID state
QP lifecycleRC/UC/UD selected an unroutable fe80:: GID and failed with ENETUNREACH
Enumerationa three-entry live GID table was expanded into 1,024 entries

Verification path