Testing
Β§1Section map
π§ͺLTP compliance scan
The full 1,869-case LTP scan: the 2026-11-15 audited round 1122/1869 (60.0%) β the official 2026-11-16 confirmation scan (scan3) corrected 1138/1869 (60.9%) β current expected 1139/1869 with the tgkill03 fix merged (x86_64+KVM). Distribution, the round trend, the isolated-rerun correction protocol and the NOT-OURS adjudication methodology.
Open βπLTP full case list
All 1,869 cases individually, grouped into 838 families with a search filter. Each row: idx / case / verdict / rc.
Open βπ‘οΈStability
Kernel suite (riscv64 1584 / x86_64 1263 all green), formal verification standing, desktop session gates (2100 s Γ3, 45 minutes zero panic), the 22.04 GNOME 5/5 stability verification, census/heal self-correction.
Open βπ₯Stress
The standing stress-fleet cluster (11 QEMUs): LTP sweeps Γ4 + randomized testing Γ2 + memstorm + fsfuzz + a control instance; root-caused classes, the watch list and the test queue.
Open ββοΈCI pipeline
make ci β one-key full regression (2026-11-10): build β suites β image β Ubuntu/OH boot regressions (marker + KERNPANIC verdicts) β lmbench/Geekbench β full LTP β the 32-class stress loop; unattended, a KERNPANIC halts with an alert.
β±οΈLTP TIMEOUT audit
Row-by-row adjudication of the TIMEOUT class under the P0 rule (2026-11-07): 13 TIMEOUT + 1 WEDGE rows re-run β 1 kernel defect (fixed), 11 test/list defects, 1 known lost-wake family, 1 slow-progress case.
Open βπRetrospective
GNOME desktop regression retrospective (2026-11-06): how 2,825 commits of silent accumulation slipped past every gate; the two root causes fixed (munmap races, page-cache poisoning) and three anti-recurrence gate changes.
Open βΒ§2Report directory
The standing reports in docs/test/ (each a dated snapshot of a measured result; raw records live outside git in run/ directories):
| Report | What it measures | Latest |
|---|---|---|
| unit-test-report.md | In-kernel guest suite (make test), per architecture | 1584/1584 riscv64 Β· 1263/1263 x86_64 Β· 2026-11-20 |
| integration-test-report.md | Multi-process unattended testing, probes, CI smoke gate | 2026-11-01 |
| ltp-report.md / ltp-tracking.md | LTP compatibility, both rigs (riscv64 TCG / x86_64 KVM) | official scan3 1138/1869 β expected 1139/1869 (x86+KVM) Β· 1028/1869 (riscv64) |
| lmbench-report.md | lmbench per-case: Rux vs Linux (same test platform) | 2026-11-06 campaign table β Performance section |
| geekbench-report.md | Geekbench 5: all 21 workloads vs Linux | MC 7504/7027 Β· SC 2465/2165 |
| stress-fleet.md | Stress/stability/randomized/fuzz cluster design + results | standing (11 QEMUs) |
| formal-verification-report.md | Kani / proptest / SPIN / Miri standing | 157 harnesses Β· 1104 proptests |
| test-coverage.md | Coverage ledger (regenerated per change) | rolling |
| gnome-regression-retrospective.md | GNOME desktop regression retrospective: root causes, gate blind spots, anti-recurrence | 2026-11-07 (adds the park-face mechanisms) β retrospective sub-page |
| ltp-timeout-audit.md | LTP TIMEOUT-class adjudication audit (P0 rule) | 2026-11-07 β TIMEOUT audit sub-page |
Β§3Test forms
- In-kernel suite (make test): 76 test files across eight domains (memory / process / fs / ipc / signal / network / driver / syscall); all assertions run inside the QEMU guest (a kernel crate cannot host host-side
#[cfg(test)]). The CI smoke gate (smoke.yml) runs the riscv64 leg on every push. - Integration cluster: multi-process unattended runs + probe matrix (integration-test-report.md); the GUI has a 23-check pixel/serial gate (
test/ubuntu-gui/verify.py) with pixel-identical cross-boot comparison (convcheck). - Formal verification: 1104 proptests binding on every push; Kani (157 harnesses) / SPIN / Miri nightly (see the Security section).
- 4,560+ test cases in the repository (in-kernel suite 1,584 + LTP passes 1,139 + the LTP inventory 1,838 = 4,561, plus probes and benchmarks).
Β§4Method and adjudication discipline
- Every number carries its test platform: architecture, accelerator, kernel profile, host-load discipline β a number without a test platform is not a result.
- Append-only reports: a new run appends a section or refreshes the standing table with its date; history is never silently rewritten.
- Flake isolation: new full-scan failures are isolated-rerun twice before being called regressions (known flake families: accept02, getrusage04, kill02, genfmod, rename/statfs).
- NOT-OURS ledger: the same musl LTP binaries run on the host Linux kernel; failures that reproduce there are recorded as test/environment issues, not kernel gaps.
- Instrument misdiagnoses are archived too: PTGUARD/FREEGUARD false positives (mapcount=β1 is a sentinel), the first PT_LEDGER double-free β negative results with receipts stay in the record.