[BUG] One-shot replay VM loses its physical QMP input connection #93

Open
opened 2026-09-25 12:09:15 +02:00 by erikinkinen · 0 comments
Owner

Summary

The existing one-shot replay negative VM test fails in its physical QMP input observer. The observer records 59 screenshots without trusted text, then reports a broken pipe. This does not establish a guest authorization failure or successful completion of the replay check.

Reproduction

Original CI 1793, integration job 23110 attempt 1, fails worker 8 while running:

bash scripts/run-scenario.sh tests/scenarios/appliance-disk-image-exact-one-shot-replay-negative.toml components.toml

The first failure appears at 07:55:28 UTC on 25 September 2026. The original workflow was not restarted. The paired original CI 1794 passes this scenario; that pass does not identify or repair the cause.

Expected and actual behavior

Expected: the physical observer and all original consent, replay and marker checks complete within their existing bounds.

Actual: the runner reports physical QMP input injector failed with status 1, followed by physical QMP input injection failed: [Errno 32] Broken pipe. All 59 retained screenshot records have empty trusted text. The available workflow log does not establish why the QMP peer closed, the guest's final state, or which process first stopped.

Environment and identity

Integration d6e9ed08b3bc645e45a88b68697425c451c1840f, original push CI 1793/API 10773, integration job 23110 attempt 1. Use that revision's original component catalog and scenario bounds. Exact failed disk, firmware and screenshot bytes are not recovered from the workflow log; screenshot hashes and paths alone are not equivalent artifacts. No introducing revision is established.

Evidence and prior checks

Original-source control-flow review — 25 September 2026: the scenario, dispatcher and disk runner match the original d6e9ed08 bytes. The physical-input failure branch follows the VM expected-status check, and the selected exact-replay policy expects (0x10 << 1) | 1, or 33. That control flow therefore implies the runner had observed its expected VM status before reporting the injector failure. This is an inference from source and the logged branch, not a recovered VM-status/serial receipt; it proves neither the required consent/replay markers nor why the QMP connection closed. Those inner artifacts and the causal failure remain required.

Six complete original CI 1793/1794 logs total 27,432,406 bytes and have zero warning candidates. Rust and Markdown succeed. Each workflow passes 490 of 491 scenarios. CI 1793 fails only this replay test; CI 1794 instead retains the known ext4 quota timeout. Worker-scoped scenario headers and summaries identify the failures independently of expected host negative-test errors.

The separate exact-selector race observer failure has different scenario evidence; no shared cause is claimed. Keep the original failures and unchanged authority, consent, deadline and cleanup gates.

Tracking

Integration PR 12, regression acceptance, phase completion.

  • Retain and correlate the complete original workflow logs.
  • Recover the original inner QEMU/input outcome and identify the connection-loss cause.
  • Correct the responsible behavior with focused regression controls.
  • Pass the original replay and related physical-input checks with unchanged authority and deadline gates.
## Summary The existing one-shot replay negative VM test fails in its physical QMP input observer. The observer records 59 screenshots without trusted text, then reports a broken pipe. This does not establish a guest authorization failure or successful completion of the replay check. ## Reproduction Original [CI 1793](https://git.erikinkinen.fi/erix/integration/actions/runs/1793), integration job 23110 attempt 1, fails worker 8 while running: ```sh bash scripts/run-scenario.sh tests/scenarios/appliance-disk-image-exact-one-shot-replay-negative.toml components.toml ``` The first failure appears at 07:55:28 UTC on 25 September 2026. The original workflow was not restarted. The paired original CI 1794 passes this scenario; that pass does not identify or repair the cause. ## Expected and actual behavior Expected: the physical observer and all original consent, replay and marker checks complete within their existing bounds. Actual: the runner reports `physical QMP input injector failed with status 1`, followed by `physical QMP input injection failed: [Errno 32] Broken pipe`. All 59 retained screenshot records have empty trusted text. The available workflow log does not establish why the QMP peer closed, the guest's final state, or which process first stopped. ## Environment and identity Integration `d6e9ed08b3bc645e45a88b68697425c451c1840f`, original push CI 1793/API 10773, integration job 23110 attempt 1. Use that revision's original component catalog and scenario bounds. Exact failed disk, firmware and screenshot bytes are not recovered from the workflow log; screenshot hashes and paths alone are not equivalent artifacts. No introducing revision is established. ## Evidence and prior checks Original-source control-flow review — 25 September 2026: the scenario, dispatcher and disk runner match the original `d6e9ed08` bytes. The physical-input failure branch follows the VM expected-status check, and the selected exact-replay policy expects `(0x10 << 1) | 1`, or 33. That control flow therefore implies the runner had observed its expected VM status before reporting the injector failure. This is an inference from source and the logged branch, not a recovered VM-status/serial receipt; it proves neither the required consent/replay markers nor why the QMP connection closed. Those inner artifacts and the causal failure remain required. Six complete original CI 1793/1794 logs total 27,432,406 bytes and have zero warning candidates. Rust and Markdown succeed. Each workflow passes 490 of 491 scenarios. CI 1793 fails only this replay test; CI 1794 instead retains the known [ext4 quota timeout](https://git.erikinkinen.fi/erix/integration/issues/18). Worker-scoped scenario headers and summaries identify the failures independently of expected host negative-test errors. The separate [exact-selector race observer failure](https://git.erikinkinen.fi/erix/integration/issues/64) has different scenario evidence; no shared cause is claimed. Keep the original failures and unchanged authority, consent, deadline and cleanup gates. ## Tracking [Integration PR 12](https://git.erikinkinen.fi/erix/integration/pulls/12), [regression acceptance](https://git.erikinkinen.fi/erix/integration/issues/6), [phase completion](https://git.erikinkinen.fi/erix/integration/issues/65). - [x] Retain and correlate the complete original workflow logs. - [ ] Recover the original inner QEMU/input outcome and identify the connection-loss cause. - [ ] Correct the responsible behavior with focused regression controls. - [ ] Pass the original replay and related physical-input checks with unchanged authority and deadline gates.
erikinkinen added this to the Alpha 1 milestone 2026-10-30 09:39:19 +01:00
Sign in to join this conversation.
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
erix/integration#93
No description provided.