I believe this fixes the root cause of the kvmclock warp.  It's quite a 
plausible phenomenon, and explains why it was so easy to produce.

Currently it depends on some other patches; I can send a whole patchset, 
but with all the patch activity, it isn't clear what has been applied 
and to what trees.  Where have Glauber's recent patches been applied?

I am looking for comments if this is a reasonably good explanation and 
fix for the problem.

I realize I messed up the overshoot calculation, it is not converted to 
nsec, but the debug stats are just for debugging.

Thanks,

Zach