Re: [RFC v6 10/11] accel: introduce AccelCPUClass extending CPUClass

[Claudio Fontana]

On 11/27/20 12:22 PM, Claudio Fontana wrote:
> On 11/27/20 9:59 AM, Claudio Fontana wrote:
>> On 11/27/20 7:21 AM, Paolo Bonzini wrote:
>>> On 26/11/20 23:32, Claudio Fontana wrote:
>>>> +    if (acc) {
>>>> +        object_class_foreach(accel_init_cpu_int_aux, cpu_type, false, 
>>>> acc);
>>>> +    }
>>>
>>> Any reason to do it for cpu_type only, rather than for all subclasses of 
>>> CPU_RESOLVING_TYPE?  This would remove the cpu_type argument to 
>>> accel_init_cpu_interfaces and accel_init_interfaces.
>>>
>>> Otherwise I haven't done a careful review yet but it looks very nice, 
>>> thanks!
>>>
>>> Paolo
>>>
>>
>> Hi Paolo,
>>
>> yes, I thought to pass cpu_type in order to set the interface only for the 
>> cpu that is actually used,
>> instead of looping over all cpu models, just to be a bit quicker, but both 
>> things should work.
>>
>> Ciao,
>>
>> Claudio
>>
> 
> Note that this actually creates a bug that is caught _ONLY_ by
> 
> acceptance-system-centos.
> 
> The gist of it is that cpu_type (or current_machine->default_cpu_type) is 
> _not_ guaranteed to be set,
> the code there is a bit misleading I think.
> 
> I'll look into it, but just wanted to warn early about it.
> 
> Ciao,
> 
> Claudio
> 

This seems to be due to "-machine none", is machine none supposed to have no 
default cpu_type?
Is it expected that for machine none current_machine->cpu_type is NULL, or is 
it a bug?

Thanks,

Claudio