Re: [Qemu-ppc] [PATCH] spapr: support CPU hotplug

[Alexander Graf]

On 27.08.2013, at 11:03, Alexey Kardashevskiy wrote:

> On 08/27/2013 06:54 PM, Alexander Graf wrote:
>> 
>> On 27.08.2013, at 09:41, Alexey Kardashevskiy wrote:
>> 
>>> On 08/27/2013 05:02 PM, Paolo Bonzini wrote:
>>>> Il 27/08/2013 08:37, Alexey Kardashevskiy ha scritto:
>>>>>>> So this is here to make sure we don't accidentally get out of halted 
>>>>>>> state by an interrupt on that vcpu. Could you please somehow make that 
>>>>>>> part obvious? Either by adding a comment or by only explicitly masking 
>>>>>>> DEC and EE and a comment :).
>>>>>>> 
>>>>>>>> +    cs->exit_request = 1;
>>>>>>> 
>>>>>>> This should probably be qemu_cpu_kick_self().
>>>>>> 
>>>>>> Uh, no, I don't think so.  This is there purely to make sure we exit
>>>>>> the inner loop, and actually test cpu_can_run() which will test
>>>>>> halted.  AFAICT qemu_cpu_kick_self() won't do anything similar.
>>>>> 
>>>>> rtas_stop_self() eventually returns to kvm_cpu_exec() which calls
>>>>> qemu_cpu_kick_self() and resets cs->exit_request before return so I do not
>>>>> really see the difference in behaviour. And actually both ways CPU stops 
>>>>> in
>>>>> exactly the same way. What do I miss?
>>>> 
>>>> What about TCG?
>>> 
>>> Oh. Right. TCG :(
>>> 
>>> qemu_cpu_kick_self() crashes the guest and cs->exit_request works fine.
>>> 
>>> Why? Both should work? What is the expected behavior here? Thanks.
>> 
>> Hrm. To me exit_request always was an internal piece of state that the inner 
>> loop uses to find out whether to exit, but not something we should randomly 
>> set from a device (and hypercalls / rtas calls are very similar to devices). 
>> So I would like to not have any code in hw/ that modifies it.
>> 
>> However, we need the functionality of breaking out of the main loop, I agree.
> 
>> Maybe what you are really looking for is
>> cpu_interrupt(CPU_INTERRUPT_HALT). That sets halted = 1 and exits the
>> main loop, because it's an interrupt.
> 
> cpu_interrupt(CPU_INTERRUPT_HALT) works fine for TCG but does not for KVM
> (the rtas call returns to the guest and it reports BUG).

How about cpu_exit()? That looks exactly like what we need.


Alex