[PATCH v28 00/17] ALSA: virtio: preserve PCM progress
From: Joonho Lee
Date: Wed Nov 11 2026 - 23:41:05 EST
This replaces v27 in full. Apply all seventeen patches directly to base
commit a4311e9447bbb219044db38897f95be112f46a8a. The previous public
version is at:
https://lore.kernel.org/linux-sound/20261112004704.49758-1-mynameisjune111@xxxxxxxxx/T/
Changes since v27 are confined to patch 12. The other patch diffs are
unchanged.
I reproduced both capture findings from the Sashiko review. If request
N+1 completed before N, RESET could leave N+1's already completed
payload eligible for later delivery. After an external XRUN, rejecting N
also left N+1 buffered until generation cleanup. PREPARE cleared that
state in the control test, so this does not establish a permanent
recovery failure.
RESET now marks valid completed payloads present at the pointer
observation as discarded while retaining their ordered placeholders.
This lets the commit sequence advance when older requests return.
Transport-owned requests and data arriving after the observation remain
eligible to commit. Repeated RESET preserves the discard, and invalid
response status still reports XRUN. Rejected completions also release
returned buffered entries without resetting the sequence of requests
still owned by the transport.
The failed native PAUSE path previously reopened the queue with the STOP
quiet boundary and capture_stopping flag still set. A running stream
could then accept an empty completion without the normal XRUN check.
Rollback now restores the prior flags under the substream lock and
checks the reset epoch before reopening admission.
ARM64 QEMU tests use the actual driver functions with test-only request
ordering and STOP allocation-failure injection. The v27 baseline fails
eight affected assertions; the candidate passes all thirteen assertions,
including normal reordering, post-observation delivery, invalid status,
and STOP empty-tail controls. The running-observation test invokes the
actual pointer-observation helper. Full RUNNING RESET ioctls are covered
separately by the production regression.
The production driver module was built separately without those hooks.
Its ARM64 W=1 build and eight QEMU regression runs passed with the
matching v27 KASAN/lockdep kernel and core modules. These cover RW/MMAP
capture, PREPARED RESET with DRAIN/reopen, RUNNING RESET, and playback
PAUSE/RELEASE, HWSYNC/poll, DROP and reopen. Capture includes 70-second
RW and resumed MMAP phases. The independent backend-sequence oracle
matched 7,104,000 frames across seven capture phases. No KASAN or
lockdep diagnostic appeared.
The kernel image and core were reused unchanged; the modified driver was
rebuilt. Fresh KCSAN and physical VZ tests were not run for this
revision. The earlier v27 KCSAN and Scarlett results describe that
earlier build. The controlled tests use synthetic requests and a known
STOP failure before submission; they do not qualify transport
timeout/uncertainty or demonstrate out-of-order completions on physical
hardware. Intermediate patch stages are application-checked; the fresh
runtime tests cover the final candidate.
Review addressed:
https://lore.kernel.org/linux-sound/sashiko-outbox-210240@xxxxxxxxxx/
base-commit: a4311e9447bbb219044db38897f95be112f46a8a