ALSA (on pipewire): continuous xruns on 0.17.3 and 0.18.1

[SludgePhD]

I'm seeing a continuous stream of xrun errors after updating from cpal 0.16 to 0.18 (and that also show up on 0.17).

These only seem to appear when requesting a non-standard sample rate (44100 instead of the native 48000), and when the requested buffer size is below 1024 (the default picked by cpal is 512). They also seem to appear at the native sampling rate, when the buffer size is not a power of two (a size of 900 starts causing them, while a size of 512 is completely fine).

These xruns do not show up in pw-top or the pipewire logs, but there is audible distortion. However, given that they persist even with a trivial data callback (data.fill(0.0)), and with --release, they cannot be caused by the callback taking too long.

These issues immediately stop appearing completely, no matter the requested stream settings, when the pipewire backend is enabled.

[roderickvd]

Thanks. That power-of-two observation is important here.

From 0.17 onwards we changed from a 4-period setup to a double-buffering setup (2 periods). That double-buffering is consistent across all hosts. On top of that, cpal guarantees callbacks a full, constant-sized period.

Put those together and the period has to line up with PipeWire's quantum, which is a power of two. When it does (512 at 48 kHz) you're fine. When it doesn't - a non-power-of-two size like 900, or 44100 resampled onto a 48000 graph where the resampler emits ~940 frames per quantum - avail never settles on a period boundary. The write cycle is then deferred to keep the buffer constant-sized, and that deferral triggers an underrun with no further periods to spare.

The way forward may be something for the v0.19 list (#1220):

• an API to configure the period count like I proposed in feat!: add StreamConfigBuilder with cross-platform configuration support #1010
• a BufferSize refactor introducing a power-of-two or "near" variant

Until then, like you found, the native PipeWire backend or snapping to powers-of-two at the native rate avoids these misalignments.