ALSA: Reasonable buffer size / where should it be implemented?

ALSA: Reasonable buffer size / where should it be implemented?

[Knut Petersen]

I use a RME Digi 96 PAD audio card, and I do have
buffer overrun/underrun problems if I use the standard
rme96c linux driver.

I need a simple recording machine. It should not fail if
e.g. cron starts updatedb or if I start a make -j 15 icecream
compile job and decide to surf the internet while I record
the digital satellite radio 48kHz stream.

There is a lot of information about xrun problems, but
somehow that information either does not help to prevent
xruns on my system, is outdated, or asks for system
changes I do not accept.

No, JACK does not help. No, I do not need low latency.
No, I don't want to switch to rt kernels. No, I don't want
to use an audio PC without X, without cron, network, etc.

The hardware provides independent 64k ringbuffers for
capture and playback, that's not more than 85msec for a
96 kHz / 2 channel / 32 bit setup or ADAT. That's simply
not enough for reliable operation.

My private solution is a rme96.c that kmallocs  4 MB
software buffers for capture and playback, data transfer
between software and hardware buffer in the interrupt
service routine. That does efficiently prevent xruns even
on a really loaded system.

But I don't know if that is the right way to go.

Wouldn't  it be better if there would be an (optional) software
buffer one layer above the hardware driver? That would
increase reliability for all audio hardware with insufficient
hardware buffers.

Is there any other audio hardware with similar small
buffers?

Has somebody already written an "extended alsa buffer"
patch? Did I miss something?

To put it into one question: How much buffer should be
provided by an alsa hardware driver module?

cu,
 Knut

Re: ALSA: Reasonable buffer size / where should it be implemented?

[Jaroslav Kysela]

On Tue, 8 Feb 2011, Knut Petersen wrote:

> I use a RME Digi 96 PAD audio card, and I do have
> buffer overrun/underrun problems if I use the standard
> rme96c linux driver.
>
> I need a simple recording machine. It should not fail if
> e.g. cron starts updatedb or if I start a make -j 15 icecream
> compile job and decide to surf the internet while I record
> the digital satellite radio 48kHz stream.

I would determine the culprit where the user space blocks. If you write 
data to HDD, the culprit is probably there, in my opinion. The current 
PC machines can handle 82ms latencies quite well.

A right behaving application should do the quick audio sample reads to a 
big buffer and queue these data to HDD in a separate thread to not block 
the audio input.

> There is a lot of information about xrun problems, but
> somehow that information either does not help to prevent
> xruns on my system, is outdated, or asks for system
> changes I do not accept.
>
> No, JACK does not help. No, I do not need low latency.
> No, I don't want to switch to rt kernels. No, I don't want
> to use an audio PC without X, without cron, network, etc.
>
> The hardware provides independent 64k ringbuffers for
> capture and playback, that's not more than 85msec for a
> 96 kHz / 2 channel / 32 bit setup or ADAT. That's simply
> not enough for reliable operation.
>
> My private solution is a rme96.c that kmallocs  4 MB
> software buffers for capture and playback, data transfer
> between software and hardware buffer in the interrupt
> service routine. That does efficiently prevent xruns even
> on a really loaded system.
>
> But I don't know if that is the right way to go.

It's probably the right way to go. I would only copy data in a tasklet not 
directly in the interrupt routine. New hardware (like HDA) have 
scatter-gather DMA buffers with almost unlimited size, so users can just 
tune the system up.

 					Jaroslav

-----
Jaroslav Kysela <perex@perex.cz>
Linux Kernel Sound Maintainer
ALSA Project, Red Hat, Inc.

Re: ALSA: Reasonable buffer size / where should it be implemented?

[Jaroslav Kysela]

On Tue, 8 Feb 2011, Knut Petersen wrote:

> I use a RME Digi 96 PAD audio card, and I do have
> buffer overrun/underrun problems if I use the standard
> rme96c linux driver.
>
> I need a simple recording machine. It should not fail if
> e.g. cron starts updatedb or if I start a make -j 15 icecream
> compile job and decide to surf the internet while I record
> the digital satellite radio 48kHz stream.

I would determine the culprit where the user space blocks. If you write 
data to HDD, the culprit is probably there, in my opinion. The current 
PC machines can handle 82ms latencies quite well.

A right behaving application should do the quick audio sample reads to a 
big buffer and queue these data to HDD in a separate thread to not block 
the audio input.

> There is a lot of information about xrun problems, but
> somehow that information either does not help to prevent
> xruns on my system, is outdated, or asks for system
> changes I do not accept.
>
> No, JACK does not help. No, I do not need low latency.
> No, I don't want to switch to rt kernels. No, I don't want
> to use an audio PC without X, without cron, network, etc.
>
> The hardware provides independent 64k ringbuffers for
> capture and playback, that's not more than 85msec for a
> 96 kHz / 2 channel / 32 bit setup or ADAT. That's simply
> not enough for reliable operation.
>
> My private solution is a rme96.c that kmallocs  4 MB
> software buffers for capture and playback, data transfer
> between software and hardware buffer in the interrupt
> service routine. That does efficiently prevent xruns even
> on a really loaded system.
>
> But I don't know if that is the right way to go.

It's probably the right way to go. I would only copy data in a tasklet not 
directly in the interrupt routine. New hardware (like HDA) have 
scatter-gather DMA buffers with almost unlimited size, so users can just 
tune the system up.

 					Jaroslav

-----
Jaroslav Kysela <perex@perex.cz>
Linux Kernel Sound Maintainer
ALSA Project, Red Hat, Inc.

Re: ALSA: Reasonable buffer size / where should it be implemented?

[Knut Petersen]

> It's probably the right way to go. I would only copy data in a tasklet
> not directly in the interrupt routine. New hardware (like HDA) have
> scatter-gather DMA buffers with almost unlimited size, so users can
> just tune the system up.
>
Well, obviously recording to /dev/null has a better chance not to
hit xrun problems ;-)

So I'll clean up my q&d rme96.c code and submit it.

I have not had a look at the sources ... is arecord "right behaving"
or which command line recording tool could be used if it is not?

cu,
 Knut