> I know so little about low level x86 hardware its not even funny; but 
> recompiling the com driver with interrupt at 4 chars rather than 8, 
> would give us some ideas about how long things are being locked down in
> the i386 port.

I have done that very thing.  I still see overruns.  And my modem line runs
at 38400.  (This is on a 486/33 with the defective external cache disabled.)

> At 57600, we're talking about 1400us to respond before characters are
> lost.  Whats blocking for that long?  TTY is > then NET and BIO, right?

Has anyone ever attempted to measure the amount of time interrupts are held
off?  If not, I've done that kind of instrumentation before (not in netbsd)
and I might be able to find enough free time to repeat the effort (assuming
anyone would do anything with the results).