snd_pcm_drain() infinite loop

[sylware]

In a non-blocking dmix-ed plugin pipeline (see below), playing audio from a 44100Hz source to a running 48000Hz hardware pcm, will go into an -EAGAIN infinite loop in snd_pcm_drain().
I did a bit of tracing and it seems snd_pcm_drain() will always return -EAGAIN because in pcm_dmix.c:snd_pcm_dmix_sync_ptr0, avail is never >= pcm->stop_threshold

[sylware]

ALSA:HW_PARAMS START------------------------------------------------------------
ACCESS: RW_NONINTERLEAVED
FORMAT: FLOAT_LE
SUBFORMAT: STD
SAMPLE_BITS: [0 0]
FRAME_BITS: [64 64]
CHANNELS: [2 2]
RATE: [44100 44100]
PERIOD_TIME: [21333 21333]
PERIOD_SIZE: [940 940]
PERIOD_BYTES: [0 0]
PERIODS: [0 0]
BUFFER_TIME: [0 0]
BUFFER_SIZE: [7526 7526]
BUFFER_BYTES: [60208 60208]
TICK_TIME: [0 0]
ALSA:HW_PARAMS END--------------------------------------------------------------
ALSA:SW_PARAMS START------------------------------------------------------------
tstamp_mode: NONE
tstamp_type: MONOTONIC_RAW
period_step: 1
avail_min: 940
start_threshold: 1
stop_threshold: 7526
silence_threshold: 0
silence_size: 0
boundary: 4236761349448794112
ALSA:SW_PARAMS END--------------------------------------------------------------
ALSA PCM DUMP START-------------------------------------------------------------
Plug PCM: Linear Integer <-> Linear Float conversion PCM (S32_LE)
Its setup is:
stream : PLAYBACK
access : RW_NONINTERLEAVED
format : FLOAT_LE
subformat : STD
channels : 2
rate : 44100
exact rate : 44100 (44100/1)
msbits : 32
buffer_size : 7526
period_size : 940
period_time : 21333
tstamp_mode : NONE
tstamp_type : MONOTONIC_RAW
period_step : 1
avail_min : 940
period_event : 1
start_threshold : 1
stop_threshold : 7526
silence_threshold: 0
silence_size : 0
boundary : 4236761349448794112
Slave: Rate conversion PCM (48000, sformat=S32_LE)
Converter: linear-interpolation
Protocol version: 10002
Its setup is:
stream : PLAYBACK
access : MMAP_INTERLEAVED
format : S32_LE
subformat : STD
channels : 2
rate : 44100
exact rate : 44100 (44100/1)
msbits : 32
buffer_size : 7526
period_size : 940
period_time : 21333
tstamp_mode : NONE
tstamp_type : MONOTONIC_RAW
period_step : 1
avail_min : 940
period_event : 1
start_threshold : 1
stop_threshold : 7526
silence_threshold: 0
silence_size : 0
boundary : 4236761349448794112
Slave: Soft volume PCM
Control: PCM Playback Volume
min_dB: -51
max_dB: 0
resolution: 256
Its setup is:
stream : PLAYBACK
access : MMAP_INTERLEAVED
format : S32_LE
subformat : STD
channels : 2
rate : 48000
exact rate : 48000 (48000/1)
msbits : 32
buffer_size : 8192
period_size : 1024
period_time : 21333
tstamp_mode : NONE
tstamp_type : MONOTONIC_RAW
period_step : 1
avail_min : 1024
period_event : 1
start_threshold : 1
stop_threshold : 8192
silence_threshold: 0
silence_size : 0
boundary : 4611686018427387904
Slave: Direct Stream Mixing PCM
Its setup is:
stream : PLAYBACK
access : MMAP_INTERLEAVED
format : S32_LE
subformat : STD
channels : 2
rate : 48000
exact rate : 48000 (48000/1)
msbits : 32
buffer_size : 8192
period_size : 1024
period_time : 21333
tstamp_mode : NONE
tstamp_type : MONOTONIC_RAW
period_step : 1
avail_min : 1024
period_event : 1
start_threshold : 1
stop_threshold : 8192
silence_threshold: 0
silence_size : 0
boundary : 4611686018427387904
Hardware PCM card 0 'HDA ATI SB' device 0 subdevice 0
Its setup is:
stream : PLAYBACK
access : MMAP_INTERLEAVED
format : S32_LE
subformat : STD
channels : 2
rate : 48000
exact rate : 48000 (48000/1)
msbits

[sylware]

in the function pcm_rate.c:snd_pcm_rate_drain, the while loop which is supposed to write the remaining audio data to the slave using snd_pcm_rate_commit_area function, does not seem to update rate->last_commit_ptr, which is then "stuck", hence the never ending writting of data to the slave.

[sylware]

I wrote a quick and dirty patch which updates last_commit_ptr: it seems it fixes the issues.
That said I have no idea how to update properly this pointer in regards of boundary (and maybe other critical values).

[perexg]

The drain is not supposed to update last_commit_ptr. Also, unless the drain is not called multiple times, the while loop takes last_commit_ptr only once. Do you call drain in the non-block mode?

[sylware]

Sorry, I did not insist enough: yes, in non-block mode. I was looking at how to update consistently the rate plugin state using some already coded functions but as far I could understand the code, I found none. Then I tried to modify the plugin state with the least amount of consistent modifications, but I could not be so sure.

[sylware]

did you see my patch on the mailing list?

[perexg]

https://mailman.alsa-project.org/pipermail/alsa-devel/2020-April/165848.html

The patch looks basically good (we can fix it later if we encounter any issues). Create PR or resend with proper description and signed-off-by line. Thanks.

[sylware]

This is an on-the-spot patch, only tested with the software showing the issue. If you don't see any obvious issues with it, I'll engage in some broader testing on top of "now-git" alsa-lib. If I am not able to break it after some time, I'll post it properly formated on the mailing list.

[sylware]

Could not break it in my usage. I posted the patch on the mailing-list.
I hope I kept the state consistent enough for the draining to reach the end in intended use cases.

[perexg]

Applied in 29041c5 . Thank you.