the missing microseconds

[Robert Morris]

It looks like a hardware interrupt on a PentiumPro 200 costs about 5 us
more than the time spent in the interrupt routine. 1.5 is the cost
of saving and restoring CPU state (this from Benjie). But there's
another 3.3 us to set (on entry) and clear (on exit) the irq's mask bit in
the 8259A interrupt controller chip. This from benchmarking Linux's
mask_and_ack_8259A() and enable_8259A_irq() routines.

The 5 us accounts for about half the time I couldn't account for in
the benchmarks.