[chriskearney]

The merge-queue run for #202 failed retry-running-teardown at 16 KiB pages, while the build, KUnit and QEMU boot passed. This fixture and the Thunderbolt driver are unchanged by #202. The control sleeps 30 ms and then assumes the stop thread has reached cancellation; delaying that thread by 100 ms reproduces an assertion failure before teardown runs.

Wait for the fixture's cancellation event under its mutex, with a two-second deadline, then check that rings and the page pool remain intact. This also removes the unlocked cancel_waiting read. No production driver or CI acceptance-rule changes.

Validation:

• All 44 controls pass at both 4 KiB and 16 KiB pages, before and after the focused mutation run.
• The repaired control passes with deliberate 100 ms and 250 ms teardown-thread delays at both page sizes; 200 repeated controls pass.
• Removing production cancellation still aborts the focused control at both page sizes. checkpatch --strict reports no findings on the code diff.

Failing merge-queue job. The CI log identifies the case and SIGABRT; the scheduling-delay reproduction above is local evidence, not a captured CI stack trace. Implementation and validation assisted by Codex.

[greptile-apps]

Confidence Score: 5/5

[Low impact] The PR appears safe to merge; it removes the scheduling assumption without weakening the teardown checks.

Summary

The PR replaces the fixed sleep in retry-running-teardown with a bounded wait for cancellation.

• The refill teardown control waits for cancellation before checking rings and the page pool.

_{Reviews (1) · Last reviewed commit: "tests: asahi: synchronize tbnet refill-w..." · Reviewed by Greptile}

[maralcbr]

@chriskearney looks right. The old control slept 30 ms and then read cancel_waiting without gate, so a slow stop thread failed the assert before teardown ran. Now cancel_delayed_work_sync() sets cancel_waiting and broadcasts cv while holding gate (tbnet-refill-fixture.h:96-106), and wait_cancel() waits for that flag under the same mutex (tbnet-refill-tests.py:59-70). So the check runs exactly when the old test meant it to: cancel has started, the paused worker is still holding up teardown, and ring_stops == 0 && pool_destroy_calls == 0.

The extra broadcast is harmless, because every other waiter on cv loops on its own condition. A missing cancel now fails at the two-second deadline, before the runner's five-second per-case timeout. The Fixes: target, 71700a1, is the commit that added this control. The other usleep controls (:94, :104-105) still use a fixed sleep. If one of them flakes, it can get the same treatment.

Approving.