This lockless memory based synchronization in csd_lock_wait just doesn't work on all smp systems because not all of them properly implement these fancy memory fencing instructions. I've run into this before trying to do lockless queueing on a range of SMP systems. About the only thing guaranteed to work is a lock assertion because this tells the processor to flush its pipeline. So this csd code is busted on some systems.
This other hang condition with the sysret just makes it show up as a deadlock in csd_lock_wait. Jeff
