[RyanTheTide]

Carries Asahi Lina's upstream fix 855c62f853cb on integration/aurora at 9cbdd0734d9e.

WorkQueue::signal() can enqueue the final completed firmware command for destruction before publishing its ring free pointer. Retain that command until a later completion or queue teardown, as in Asahi's implementation. This replaces the earlier implementation in this PR with the upstream cherry-pick, preserving Asahi Lina's authorship, author date and original message.

The only adaptation is using this tree's workqueue::system_dfl() API at both cleanup enqueue sites. No kernel parameters or module flags are added.

Validation: git diff --check passes; strict checkpatch reports 0 errors, 0 warnings and 0 checks. The universal 7.3.0-aurora-g13w build passes with the same generated config as the q423 reference.

Runtime acceptance failed. The frozen 41-case reproducer, requested for 50 repetitions at eight jobs with stock Mesa 26.1.3, hit another GPU firmware stop during repetition 4. The first three repetitions completed; the failing chunk returned 137 with zero OOM kills, and the watchdog stopped the campaign. This upstream carry is insufficient to resolve the observed G13 stop. Full73 was canceled after this failure. Keep this PR draft pending further investigation.

Assisted-by: LLM

[RyanTheTide]

Closing this draft: the carried last-completed-work-item retention did not eliminate the G13 firmware stop in verification.

Upstream Asahi has also reverted the equivalent change in 83526ea43d4f, “drm/asahi: workqueue: Drop completed work items”. Retaining the last item pins a BufferManager slot and limits userspace to at most 127 queues. This carry should not merge; any subsequent lifetime fix needs a bounded retention mechanism that avoids pinning those slots.