Ok, here's try number 4.  This one fixes the race condition that 
Dipankar pointed out, and modifies the release_nmi() call to block until 
the rcu finishes.

I have noticed that the rcu callback can be delayed a long time, 
sometimes up to 3 seconds.  That seems odd.  I have verified that the 
delay happens there.

Also, does the __wait_for_event() code in release_nmi() need any memory 
barriers?

Also, the code in traps.c is pretty generic.  Perhaps it could be moved 
to a common place?  Or is that worth it?

Thanks again,

-Corey