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:

>> CPU A call smp_call_function_many(wait=0) that want CPU B to call
>> a specific function, after smp_call_function_many() return, we let
>> CPU A offline immediately. Unfortunately, if CPU B receives this
>> IPI interrupt after CPU A down, it will crash like above description.
> 
> How can cpu B receive the IPI interrupt after cpu A is down?
> 
> As part of the cpu A going down, we first do the stop machine. i.e.,
> schedule the stop machine worker threads on each cpu. So, by the time
> all the worker threads on all the cpu's get scheduled and synchronized,
> ipi on B should get delivered.
> 

Actually, those two examples have the same reason, that is how long
the destination CPU will receive the IPI interruption?

If the stop machine threads can schedule in CPU B during the IPI
interruption delivering, It will occur those issue.

I understand what you say but let me confuse is how we ensure it? The IPI
interruption is delivered over the APIC bus, It need several CPU instruction
cycle I guess, I also have read the spec of Intel 64 and IA-32, but not find
the answer, could you point out for me?

Thanks,
Xiao