Re: [Qemu-devel] [RFC PATCH] target-ppc: Add compatibility between P7/P7+ and P8E/P8

[Alexey Kardashevskiy <aik@ozlabs.ru>]

On 06/28/2014 10:31 AM, Alexey Kardashevskiy wrote:
> On 06/28/2014 10:00 AM, Alexey Kardashevskiy wrote:
>> On 06/28/2014 02:14 AM, Alexander Graf wrote:
>>>
>>> On 27.06.14 17:54, Alexey Kardashevskiy wrote:
>>>> At the moment POWER7+ and POWER7 CPUs are different incompatible
>>>> families in QOM. The same is valid for POWER8E and POWER8 CPUs.
>>>> However, these couples are architecturally equal and there is no
>>>> good reason, for example, not to let run -cpu POWER7 on the real
>>>> POWER7+ CPU machine.
>>>>
>>>> This introduces one more level in hierarchy of POWERPC CPU classes.
>>>> New macro POWERPC_FAMILY_2 takes a family class and the parent family
>>>> class and, for example, for POWER7+ the hierarchy looks like:
>>>> TYPE_CPU
>>>> TYPE_POWERPC_CPU
>>>> POWER7-powerpc64-cpu
>>>> POWER7+-powerpc64-cpu
>>>>
>>>> This registers new dynamic POWERPC CPU classes for all classes between
>>>> the lowest one which matches the real PVR and TYPE_POWERPC_CPU.
>>>> So for POWER7, it is still going to be just a single dynamic "POWER7"
>>>> class but for POWER7+ inherited from POWER7 there are going to be
>>>> 2 dynamic classes  - "POWER7+" and "POWER7" so management software
>>>> can use both to ensure successful migration.
>>>>
>>>> Since POWER7+ inherits from POWER7 and POWER8E from POWER8, this
>>>> removes recurring pieces of code. CPUs with shorter names were chosen
>>>> as parents.
>>>>
>>>> Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
>>>> ---
>>>>
>>>> This is rather RFC patch and there is no hurry in reviewing this,
>>>> and this is not 2.1 material and everyhting, just tried to solve
>>>> a QOM puzzle here :)
>>>
>>> I'm not sure - I'd rather make sure we have this sorted out for 2.1 so we
>>> can keep the -cpu list stable.
>>>
>>> Could we make the PVR matching a function callback rather than value+mask?
>>> Then we could have p7 and p8 just match on 2 different PVR ranges.
>>
>> POWER8:
>>>>> bin(0x4d)
>> '0b1001101'
>>>>> bin(0x4b)
>> '0b1001011'
>>>>> bin(0x4c)
>> '0b1001100'
>>
>> 4D == POWER8, 4B == POWER8E, 4C - does not exist and when it will, I do not
>> know what it is going to be.
>>
>> POWER7:
>>>>> bin(0x3f)
>> '0b111111'
>>>>> bin(0x4a)
>> '0b1001010'
>>>>>
>>
>>
>> What should mask look like for P7 and P8?
> 
> Re-read your email :) Callback you say... And we will loose POWER7+ and
> POWER8E. May it is all right as we drop bottom 16bits already...


Ufff. Ok, I tried adding a callback:
int (*pvr_match)(uint32_t pvr);

Ok. Then POWER7 has fw_name = "PowerPC,POWER7" and POWER7+ has fw_name =
"PowerPC,POWER7+". Hosts use the same names, with and without "+", QEMU
uses them to make up CPU nodes in /proc/device-tree/cpus/. If it is going
to be one CPU class, won't we need yet another callback for these fw_name's?

And this fw_path is just a field (or a typename is used if fw_path==NULL)
and since it is a DeviceState's property and used directly, I cannot
intercept this so I need to hack the original class somehow.

Make it
int (*pvr_match_and_init_fw_path_if_matched)(uint32_t pvr);
?


-- 
Alexey