Since commit fe981e6756 ("ALSA: virtio: use ack callback"), an I/O
message is only queued to the device once the application has filled
the whole period: virtsnd_pcm_msg_send() accumulates the transferred
bytes in msg->length and queues the message when it reaches
period_bytes. msg->length is reset to 0 only when the device completes
the message.
If the substream is stopped (e.g. snd_pcm_drop()) while a period is
only partially filled, the corresponding message is never queued and
keeps its partial length. virtsnd_pcm_prepare() resets msg_count,
hw_ptr and the indirect PCM state, but not msg->length. After the
next start, the stale length is added to the new data, the sum never
equals period_bytes, and the message is never sent again. Playback
then stalls once hw_ptr reaches that period: writes fail with -EIO
("DMA or IRQ trouble") and drain fails with -EIO. Capture is affected the
same way after a partial read.
Clear the fill level of all messages in prepare. Outside of the resume
path, virtsnd_pcm_prepare() already refuses to run while I/O messages
are pending, so no message is owned by the device at this point and the
reset does not race with message completion.
Reproduced in QEMU (virtio-sound-pci, wav audiodev): prepare, write 1.5
periods, drop, prepare, write 16 full periods, drain. Without this
patch WRITEI returns -EIO after 14 periods and DRAIN fails; with it all
16 periods are played and DRAIN succeeds. Period-aligned prefill works
in both cases. The same sequence on capture (read 1.5 periods, drop,
prepare, read 16 periods) fails with -EIO after 10 periods without this
patch and reads all 16 with it.
Found by code inspection, confirmed with the reproducer above.
Fixes: fe981e6756 ("ALSA: virtio: use ack callback")
Assisted-by: LLM
Signed-off-by: Ting-Han Hou <ue081723@gmail.com>
Link: https://patch.msgid.link/20261011125426.35745-1-ue081723@gmail.com
Signed-off-by: Takashi Iwai <tiwai@suse.de>
f0e9db0f84