Hi Takashi, This v14 series replaces v13 and applies as twelve patches to sound.git for-next at a4311e9447bbb219044db38897f95be112f46a8a. The patch order places the ALSA PCM synchronization prerequisites before the virtio ACK and teardown changes, so the driver can keep ACK admission closed while the core resets the PCM pointer. Targeting 7.5 or later works for me; I have no urgent deadline.
The changes address the 16 raw Sashiko findings from v13; several findings share a fix and its validation. They close ACK admission before teardown waits while keeping prepared runtime storage valid until device-owned capture requests drain. They prevent event delivery before ALSA objects are ready, preserve XRUN state through recovery, and keep capture accounting consistent across RESET, short RX completion, sub-byte ACKs and NOMMU post-STOP copies. Failed-probe cleanup reclaims event buffers without dispatching them into uninitialized ALSA state. Playback RESET returns -EBUSY while TX requests remain device-owned and leaves the pointers unchanged. DROP/PREPARE retires that generation before retrying. Linked PCM RESET retains its first-error and partial-completion behavior. Validation includes 21 selected QEMU runs with expected fixture exit codes, module hashes, unchanged inputs, VM cleanup and decoded kernel diagnostics checked. The callback boundary suite passed 24 cases; ACK/START failure to DROP recovery passed 18 cases; the forced ACK/teardown race passed 12 cases under KASAN/lockdep and 12 under KCSAN_STRICT. Fourteen ordinary and staged capture segments, totaling 14,208,000 frames, matched the independent sample oracle. Separately, all twelve ordered v14 stages built, with the recorded source hashes matching each stage. Four targeted intermediate-stage QEMU runs matched their expected exits, module hashes, unchanged-input checks, VM cleanup and quiet decoded KASAN/lockdep logs. The linked-peer checks accepted baseline ACK 3/3 and rejected candidate ACK with -EBADFD 3/3; three control cases per version and six candidate recovery cases passed. The stage-6 event probe dispatched one enabled event, blocked pre-init dispatch and passed rebind I/O. The stage-7 short-RX case published zero progress, latched XRUN and reopened for 96,000 bytes. Capture-stage host checks passed 15 assertions across stages 7–9, including the teardown gate's prepared-state lifetime. Two physical VZ runs used a Samson Go Mic at 48 kHz mono. Each run included a 70-second capture, three five-second reopen captures, RUNNING and PAUSED RESET, and normal shutdown. All recordings had nonzero input and zero XRUN errors. These VZ runs used the experimental capture compatibility flags. The NOMMU post-STOP path passed 22 assertions in an extracted-function host shim; no CONFIG_MMU=n kernel or device run was performed. Event admission used synthetic events; physical system suspend was not exercised. Docker Desktop, other input devices, and broad load or format coverage remain outside these results. LLM assistance was used for review analysis, implementation, fixtures and this explanation. The patches carry Assisted-by: LLM. Previous series: https://lore.kernel.org/linux-sound/[email protected]/T/ Thanks, Joonho base-commit: a4311e9447bbb219044db38897f95be112f46a8a
