Re: [Alsa-devel] OSS driver removal, 2nd round (v2)

On Tue, 11 Jul 2006, Adam Tlałka wrote:

On Mon, 10 Jul 2006 20:39:03 -0400
Lee Revell <rlrevell@xxxxxxxxxxx> wrote:

On Tue, 2006-07-11 at 01:59 +0200, Olivier Galibert wrote:
ALSA lib has something like 7 different methods just to play a sound.
Their view of "low level" is quite interesting. Using it is pure
hell. Debugging what you've done is worse. And don't bother to hope
that your code will still work in six months.


A small FAQ:

Q: But OSS is kewl and ALSA sucks!
A: The decision for the OSS->ALSA move was four years ago.
If ALSA sucks, please help to improve ALSA.

The problem is that ALSA is done a Windows way with too many not always
working ways of obtaining sound and lack of user docs. So there is still

The docs are available:
http://www.alsa-project.org/alsa-doc/alsa-lib/pcm.html

If you have some comment or idea to improve it, please, post it to
the relevant mailing list.

the general question: Is it the proper way? How I can help improve ALSA
if I just dislike its design?

Then it is your problem. I've not seen a single problematic technical
information from your mails.

I spent some time improving dmix and aoss in the past. I have a version
of aoss which uses callbacks and works properly with some not properly
written OSS apps. It's very interesting that with some apps aoss method
gives better sound A/V synchronization then using native ALSA app
support ;-). But callbacks not always work and LD_PRELOAD method is
generally not secure and some kind of a hack. So I just don't think it
could be improved with it's current design. In my opinion it should work
out of the box without need to rewrite old or proprietary apps and
messing with LD_PRELOAD. So /dev/dsp compatibility is a must. But with
all input/output mixing of course. Userspace thread method leads to many
problems - swapping and rescheduling leads to bad acustic effects. Of
course if an app can't supply data when it should we can't do anything
but if scheduling or swapping causes problems the design seems to be
bad.

There are many devices which need to be served at the some critical
point in time
- sound cards, video and tv grabbers, some dedicated lab cards, i/o
ports etc. Generally on the really low level this is a kernel driver
which sticks to the hardware and should ``know'' how critical is to
serve hw requests just in time. If we have data in a buffer programing
DMA transfer is a quite quick operation but must be done at the proper
moment. So maybe we should delay a bit other io for example and just
do it. But it must be done in kernel because kernel should know these
critical time parameters for all devices in the system and decide
which to serve and when. Maybe this is not needed for all devices but
for audio/video devices it's a must. Better kernel support for these
kind of devices is absolutely needed.

You're a bit mixing things:

a) we're not trying to be more compatible than OSS code in kernel, if you
like to do the mixing in kernel, simply write a new ALSA lowlevel
driver which will do it; I'm sure when the quality of your code will be
good, we'll include it to the ALSA tree, but we are not going this
way unless someone else will maintain this code
b) you're requestion to avoid scheduler participation in the audio
processing and yes, we have this solution already (dmix), but not in
kernel
c) as you said, the kernel should contain only critical code to drive
hardware, sample rate conversion, sample format conversions and so on
are NOT part of this code in my opinion

ALSA in lib way has its limitations and drawbacks and adding more
feaures this way leads to more complications only IMHO.

Which limitations? We can do all things like OSS API. The whole point of
all problems is that OSS has the API entry point in syscalls which is
quite bad so the redirection is problematic.

Jaroslav

-----
Jaroslav Kysela <perex@xxxxxxx>
Linux Kernel Sound Maintainer
ALSA Project, SUSE Labs

Relevant Pages