[BUG?] Setting (start_threshold > stop_threshold) makes snd_pcm_writei(a_small_buffer) XRUN immediately

[Geraldo Nascimento]

On Mon, Oct 04, 2021 at 05:48:49PM +0800, Zhang Boyang wrote:
> On 2021/10/2 12:45, Geraldo Nascimento wrote:
> > On Fri, Oct 01, 2021 at 10:01:34PM -0300, Geraldo Nascimento wrote:
> >> On Fri, Oct 01, 2021 at 08:33:23PM -0300, Geraldo Nascimento wrote:
> >>> On Sat, Oct 02, 2021 at 01:31:12AM +0800, Zhang Boyang wrote:
> >>>> Hello,
> >>>
> >>> Hello, Zhang!
> >>>
> >>>>
> >>>>     I'm using ALSA to develop an audio-streaming application. I try to
> >>>> use start_threshold and stop_threshold in combination with small
> >>>> buffers. However, I think I probably found a bug about this.
> >>>>     I'm setting start_threshold=100 and stop_threshold=50. I'm also using
> >>>> a buffer of 44 frames. When I call
> >>>> snd_pcm_writei(the_small_44_frames_buffer), pcm state came to XRUN from
> >>>> PREPARED directly. I think this is a bug because the stream hasn't
> >>>> started. It's hard to say a xrun occurred while stream not started.
> >>>>     I'm wondering if this is a ALSA-bug or a misuse of ALSA. A simple bug
> >>>> test program is attached.
> >>>
> >>> No, I don't think it's a bug. You're bound to run into problems with a
> >>> period size of only 44 frames.
> >>>
> >>> Moreover, working with the code you provided, I was able to get a RUNNING
> >>> state without XRUNs with a period size of 4410 frames (100 milliseconds of
> >>> audio) but I had to comment out snd_pcm_sw_params_set_stop_threshold() for
> >>> it to work or I'd have those instant XRUNs.
> >>>
> >>> That's how snd_pcm_sw_params_set_stop_threshold() is supposed to work by
> >>> the way. It creates a XRUN once the threshold is hit.
> >>
> >> Oh, and snd_pcm_sw_params_set_stop_threshold() is a little
> >> counter-intuitive. The threshold you give the function is compared to
> >> the available free space in the ring buffer, not the filled space.
> >>
> >> That's why lowering the stop threshold makes XRUNs more likely...
> >>
> >> Thanks,
> >> Geraldo Nascimento
> > 
> > Zhang,
> > 
> > I was able to make your code work in the end, even with 44 frames
> > writes, which is unadvisable since it eats too much CPU specially if
> > they are low-end like the ARM64 box I'm writing you from.
> > 
> > The trick was to use high values such as 500000. Lower values such as
> > 250000 were enough to trigger XRUNs in my machine. Of course you can set
> > it to -1 and it will ignore XRUNs but that's a bit like setting it to
> > 16777216 or 4096 * 4096 which is a hardcoded limit in alsa-lib, I
> > believe.
> > 
> > And plainly ignoring XRUNs may sound bad under load, believe me.
> > 
> > Thank you,
> > Geraldo Nascimento
> > 
> > 
> 
> Hi,

Hey there, Zhang,

> 
> Thank you for your reply.

It was my pleasure.

> 
> I realized my mistake about the meaning of stop_threshold. The 
> stop_threshold should be (bufsz-50), where (bufsz = alsa buffer size). I 
> fixed my code and new code is attached.
> 
> However, the problem still exists. I'm trying to minimize audio latency, 
> so I must use a small buffer. In fact, the problem is not related to the 
> absoulte size of buffer, but the relationship between the size of 
> start_threshold, stop_threshold and the buffer. The problem still exists 
> even using manual start mode (i.e. start_threshold>bufsz).

Sure it's the relationship.

If you insist on using (bufsz-50) you will get those instant XRUNs
forever.

Try using ((bufsz - frames) + 1) and enlarge your start_threshold
to at least 264. Your original code works for me with such values but
your mileage may vary. If you get XRUNs, play with the threshold values,
enlarging stop threshold and/or enlarging start threshold.

> 
> Currently, I'm using a workaround. After PREPARED or XRUN'd, I manually 
> buffer incoming audio frames, and when (n_buffered_frames >= 
> start_threshold && n_buffered_frames >= bufsz-stop_threshold), I write 
> them in a single write() call. Note that the write must in whole, any 
> short write may result in immediate XRUN.

Clever. Reminds me of the workings of the new low latency mode recently
introduced for snd-usb-audio. I haven't been able to test your revised
code but that's certainly something you'd like to keep in your codebase,
to deal nicely with eventual XRUNs.

Thank you,
Geraldo Nascimento

> 
> Anyway, Thank you for your advice. Have a nice day. :)
> 
> Zhang Boyang