Had this around for a while too. Quite a bit of arch work to get through here, and I haven't done most of it (I probably butchered ppc64, too).
It does seem to work OK on i386 and ia64, though. I'd like to get confirmation that my approach is correct before putting more work into it.
It can save locked memory ops and IPIs during scheduling operations. I don't have access to any big systems that might most beneift from the improved scalability, but it is worth a % or so on a dual Xeon system here (on schedule happy workload).
But am I horribly misinformed? ;)
