Hm, wouldn't it just work? In ppc6xx_idle() the _TLF_NAPPING bit is set. 
If as a result of mtmsr only the EE bit is set and we get an interrupt, we 
end up in transfer_to_handler(), check the flag, it is set, so we branch 
to power_save_6xx_restore(). There we clear NAP/DOZE and just jump to 
transfer_to_handler_cont(). Why did you say we'd miss the check for 
need_resched (on IRC)? How is this case difference from if we really did 
go to NAP / DOZE?

Are there other places in the kernel where we rely on setting MSR:POW and 
some other bit atomically?

Indeed, the napping "recovery" code might save us here... funny, as it
wasn't implemented to handle that case at all...

Ben.