Hehe, I recently did something similar (in Python, though) ...

Good to now i'm not alone playing with SCOs !!

The thing basically works... exect after 1or 2 minutes playback, the 
program suddenly exits.
Then i just have a few seconds left and *bang* . Kernel Panic :-(


... and had similar results. Just look at the memory consumption when
this happens. Apparently all incoming SCO data is buffered in kernel
memory for you to read, so you *must* read it, even if you only ever
intend to send data.

I'd also suggest you simply use select() to wait for incoming data and
then only send something when you received (and read!) something. This
nicely gives the right timing without guessing the delay parameter.

Quite interesting is the way you solved the issue :-)
I'm gonna give a try to this select way of things :-)
By the way i just ran another test using hsplay. I basically send a SCO packet every second,
which is obviously not enough to sustain the SCO packet flow.
Using hcidump i am able to see the SCO flow coming from the Headset.
... and no crash at all :-)

Which for me basically means that the bug is somewhere on the PC-->Heaset flow, where something must be handled the wrong way...

Cheers,

Fabien