Fix N class SMP

The main fixes are:

 - memory barriers to our space and control register moves
 - fix for the N class merced bus problem which can't have more than
   one PxTLB broadcast outstanding at once
 - make smp_call_function() wait until the function completes

This now boots and runs on a 32MB N4000 in Fort Collins with 2 cpus
and discontig mem support.
Committed-by: James Bottomley <jejb@parisc-linux.org>
Committed-by: Randolph Chung <tausq@parisc-linux.org>