Hello Tejun Heo,

I've backported the  schedule_on_each_cpu()  "direct excution" patch on 
3.0.30-rt50,
and  It  fixed my problem.

attachment is the effective patch.

However, I do not understand why  machine1 can expose problem,  but 
machine2 not.

I guess, because it's  rt-kernel's  preempt level related, so , is this 
difference due to cpu performance?

How do you think about this ?

Thank you~