Jesse Barnes wrote:

>Looks like a bit more context has changed.  Manfred, care to respin against 
>-mm3 so I can test?
>
>  
>
The patches are attached. Just boot-tested on a single-cpu system.

Three  changes:
- I've placed the per-group structure into rcu_state. That's simpler but 
wrong: the state should be allocated from node-local memory, not a big 
global array.
- I found a bug/race in the cpu_offline path: When the last cpu of a 
group goes offline then the group must be forced into quiescent state. 
The "&& (!forced)" was missing.
- I've removed the spin_unlock_wait(). It was intended to synchronize 
cpu_online_mask changes with the calculation of ->outstanding. Paul 
convinced me that this is not necessary.

--
    Manfred