Re: + generic-ipi-fix-the-race-between-generic_smp_call_function_-and-hotplug_cfd.patch added to -mm tree

[Xiao Guangrong <xiaoguangrong@cn.fujitsu.com>]

Suresh Siddha wrote:
> On Sun, 2009-09-20 at 19:55 -0700, Xiao Guangrong wrote:
>> Suresh Siddha wrote:
>>> On Wed, 2009-09-16 at 20:00 -0700, Xiao Guangrong wrote: 
>>>> How about manual check/handle pending IPI interruption in the CPU context?
>>>> like this:
>>>> --- a/kernel/cpu.c
>>>> +++ b/kernel/cpu.c
>>>> @@ -173,6 +173,9 @@ static int __ref take_cpu_down(void *_param)
>>>>  	struct take_cpu_down_param *param = _param;
>>>>  	int err;
>>>>  
>>>> +	generic_smp_call_function_interrupt();
>>>> +	generic_smp_call_function_single_interrupt();
>>>> +
>>> At this place, how will you ensure that the smp_call_function initiated
>>> by this dying cpu has reached and got serviced at its destination?
>>>
>> Suresh, sorry for my poor English, Do you mean that how we ensure it has
>> pending IPI request in the dying cpu?
>>
>> generic_smp_call_function_*() will check it, if the cpu has pending request,
>> then handle it, else directly return.
>>
> 
> I am referring to the missing csd_lock_wait() here that you had in the
> first version of your patch. Let's say, if cpu X is going offline, we
> need to ensure that the smp_call_function() initiated by cpu X (i.e.,
> smp_call_function IPI sent to some other cpu's from cpu X) got serviced
> before cpu X goes offline. We can't do csd_lock_wait() here, as that
> might deadlock (as all the other cpu's are already in stop machine with
> interrupts disabled).
> 

It not happen because the preemption is disabled while send IPI request and
can't schedule to stop machine path, it also stop cpu down.

Thanks,
Xiao