I tried a bundh of scenarios on some computers I had at work.  Nothing I
did would keep my 400 MHz system from losing characters on the serial
driver if I ran my cdrom full tilt and used > 38400 buad.  At or under
38400, turning on the fifo was enough, though I still saw one or two
overruns over the course of 15 minutes.  There's clearly something about
the cdrom driver that's staying at splhi way too long or causing back to
back interrupts that keep the serial interrupts from sneaking in.  At 9600
baud down (which is where the mouse should live), turning on the fifo
worked 100% of the time.

The serial driver (at least from interrupt to mouse stream interpretation)
is not  fundamentally different now than in the 3rd edition.  If the same
mouse hardware works with the 3rd edition and loses chars with the 4th I'm
moderately flabergasted.

For straight (non-mouse) serial ports, there is one difference: the 3rd ed
had much smaller queues and turned off RTS fairly often, effectively reducing
the port throughput to a fraction of what it was really set to.  It might be
that this was avoiding overruns.  I need to figure out a way to do the same
thing when the computer really can't keep up.  That should get rid of the
overruns for things like PPP.  Won't effect mice since I believe that they
ignore hardware flow control, at least the ones I have do.

The current aux/mouse on sources turns on the fifo.