I don't pretend to understand alpha memory ordering.. but I do have a
concern, looking more from the MI level:

On systems with "interesting" memory coherency models, the appropriate
barrier instructions should be included as part of __cpu_simple_lock()
and __cpu_simple_unlock()... if the effects of writes can appear to be
reordered around lock and unlock, we're doomed.

A more interesting question is when a mutex is used only to serialize
updates to locations (such as a PTE) which are read without holding a
lock..  if we can't allow this for this in a simple way, we're going
to have to be a lot more paranoid about what we allow in MI code.

					- Bill