after looking at it some more, it seems like we can get away without
changing the interrupt level of statclock().  another way to look at
the problem is that we already have a two-level interrupt scheme for
the scheduler (the troublesome part of statclock()) on the sparc via
the two different clock interrupts.  the bulk of the scheduler only
needs to run at "splsched()" (which is documented in the spl manpage,
defined on only the m68k-based ports, and not used at all).
I propose to resurrect splsched() and use it to solve
the zs overrun problem.  diffs based on this morning's sup are at

	http://www.chuq.com/netbsd/splsched.diff

and a GENERIC kernel with above diffs applied is at

	http://www.chuq.com/netbsd/netbsd.splsched.gz

(diffs only complete for sparc currently, other ports need merely
add a definition for splsched(), or possibly correct an existing definition.)
these changes eliminate zs fifo overruns on my sparc1 and IPX at 38400 baud.

there's still a slight bit of funniness after these patches:
p_usrpri, p_priority and curpriority are still accessed outside
splstatclock() in arch/sparc/sparc/trap.c:userret(),
but it's been like that all along so I left it alone for now.
I don't know exactly what the different between these fields is anyway,
can someone explain?

-Chuck