Uros Bizjak wrote:

>   Could you put a printk() in snd_sb16_*_pointer() functions
> (sb16_main.c souce)? This printk() should print dma channel and value of
> ptr, so we can see when/if DMA channels gets stuck.

Done.  The result put 2.6 megabytes into /var/log/messages.
The head -10 and tail -50 of the relevant section is
attached below;  if anybody wants more detail the whole
shebang is available at
   http://www.av8n.net/~jsd/alsa/messages.txt

>   BTW: It looks like another sb16 hardware bug to me. On module re-load,
> drivers reset sb16 chip, that is why module reload helps.

Plausible....  But note that the cs4239 module gets i/o
errors when provoked by the same program, so there may
be driver design weaknesses involved, too.

> Perhaps you can derive a small test case from your application, which
> will help developers to stresstest soundcard and perhaps PCM layer?

My program is not super-small, but the relevant
passages are simple and easy to find.
You can help yourself at
   http://www.monmouth.com/~jsd/turbid/

Compile the thing and invoke it as
   ./turbid -F 440
or
   ./turbid -v -F 440

Note that just for you I added the --snooze <tt>
option, which will cause it to generate an
overrun every <tt> seconds (default .5) PROVIDED
<tt> is not too small.