Re: [PATCH v14 18/22] accel: introduce AccelCPUClass extending CPUClass

[Claudio Fontana]

On 1/29/21 1:13 AM, Richard Henderson wrote:
> On 1/28/21 6:29 AM, Philippe Mathieu-Daudé wrote:
>> On 1/28/21 5:08 PM, Alex Bennée wrote:
>>>
>>> Claudio Fontana <cfontana@suse.de> writes:
>>>
>>>> On 1/28/21 2:03 PM, Philippe Mathieu-Daudé wrote:
>>>>> On 1/28/21 10:28 AM, Claudio Fontana wrote:
>>> <snip>
>>>>>> +
>>>>>> +#define TYPE_ACCEL_CPU "accel-" CPU_RESOLVING_TYPE
>>>>>> +#define ACCEL_CPU_NAME(name) (name "-" TYPE_ACCEL_CPU)
>>>>>> +typedef struct AccelCPUClass AccelCPUClass;
>>>>>> +DECLARE_CLASS_CHECKERS(AccelCPUClass, ACCEL_CPU, TYPE_ACCEL_CPU)
>>>>>> +
>>>>>> +typedef struct AccelCPUClass {
>>>>>> +    /*< private >*/
>>>>>> +    ObjectClass parent_class;
>>>>>> +    /*< public >*/
>>>>>> +
>>>>>> +    void (*cpu_class_init)(CPUClass *cc);
>>>>>> +    void (*cpu_instance_init)(CPUState *cpu);
>>>>>> +    void (*cpu_realizefn)(CPUState *cpu, Error **errp);
>>>>>
>>>>> If we want callers to check errp, better have the prototype return
>>>>> a boolean.
>>>>
>>>> Good point, the whole errp thing is worth revisiting in the series,
>>>> there are many cases (which are basically reproduced in the refactoring 
>>>> from existing code),
>>>> where errp is passed but is really unused.
>>>>
>>>> I am constantly internally debating whether to remove the parameter 
>>>> altogether, or to keep it in there.
>>>>
>>>> What would you suggest?
>>>
>>> I think it really depends on if we can expect the realizefn to usefully
>>> return an error message that can be read and understood by the user. I
>>> guess this comes down to how much user config is going to be checked at
>>> the point we realize the CPU?
>>
>> cpu_realize() is were various feature checks are, isn't it?
>>
>>   -cpu mycpu,feat1=on,feat2=off
>>   CPU 'mycpu' can not disable feature 'feat2' because of REASON.
> 
> Yes.  And while changing the return type of realize is probably a good idea, 
> it
> should be a separate patch.
> 
> 
> r~
> 

To summarize:

1) accel->cpu_realizefn extends the current cpu target-specific realize 
functions with accelerator-specific code,
   which currently does not make use of errp at all (thus, the temptation to 
remove errp from the interface until it is actually needed by a target).

2) Regarding the target-specific cpu realize functions themselves, registered 
with the pattern:

   device_class_set_parent_realize(dc, x86_cpu_realizefn, &xcc->parent_realize);
   device_class_set_parent_realize(dc, arm_cpu_realizefn, &acc->parent_realize);

   ... etc ...

   these currently all return void, and the code that realizes devices (f.e. in 
hw/core/qdev.c)
   calls these callbacks like this:

   static void device_set_realized(...)
   {
   ...
        if (dc->realize) {
            dc->realize(dev, &local_err);
            if (local_err != NULL) {
                goto fail;
            }
        }

   Ie it does not expect a bool return type for realize().

   So making realize return bool means changing all the realize functions for 
all devices in my view,
   which I don't think should be crammed in here..

Wdty?

Thanks,

Claudio