> Someone suggested profiling the zs driver.  I'd have to

I don't think this is going to tell you much. It would be more
interesting to know which pieces of code block out the serial
interrupt for too long (most likely at splhigh or splstatclk, unless
you're playing audio) by recording stack traces when the zs fifo
overflows.

Someone already did this a long time ago, and the results seemed to point
at the scheduling code.

Unfortunately (for you, not for me:-), I never see a single FIFO overflow
on my Classic running a PPP connection at 38400baud..

-pk