3.4: nohz load accounting still has some problem

Hi Peter,
With 3.4 we still has problems with nohz load accounting. In one of our
product system, the number of running processes is around 16, but loadavg is only around 8-10. Without your fix c308b56b5, it is only less than 1... Your patch does work, but not good enough. :(

After some investigation, it seems that we still have a hole, but we and not sure and haven't figure out how to resolve it. So maybe you have a good idea of whether our analysis is good and how to fix it.

So in general after your fix c308b56b5, we will fold the nohz remainder to the global one after all the cpu has calculated the real value. But there does exist some case: See
cpu 0 1 2
calc calc

idle and update
calc_load_tasks_idle

calc

Now when cpu2 calculates load, it will use calc_load_tasks_idle which has been changed by cpu1. So the load isn't accurate any more. Am I missing something here?

Thanks
Tao