The best approach of all (pie-in-the-sky) is to have all the
hardware interrupt handlers do the minimum possible, and then
schedule a kernel thread to do the remaining work.  That way
the relative priorities of the kernel threads can determine
the actual order of "mid-level" interrupt service.

Of course, this would require a nice system of preemptable
kernel threads upon which the traditional process model is
layered. (which starts to look like mach...)

If wishes were horses, beggars would ride...

Gordon