Re: [Qemu-devel] [PATCH v3 30/30] target/s390x: update maximum TCG model to z800

[David Hildenbrand]

On 02.06.2017 10:09, Thomas Huth wrote:
> On 01.06.2017 21:17, Aurelien Jarno wrote:
>> On 2017-06-01 11:04, David Hildenbrand wrote:
>>> On 01.06.2017 10:38, David Hildenbrand wrote:
>>>> On 01.06.2017 00:01, Aurelien Jarno wrote:
>>>>> At the same time fix the TCG version of get_max_cpu_model to return the
>>>>> maximum model like on KVM. Remove the ETF2 and long-displacement
>>>>
>>>> I don't understand the part
>>>> "fix the TCG version of get_max_cpu_model to return the maximum model
>>>> like on KVM".
>>>>
>>>> Can you elaborate?
>>>>
>>>>> facilities from the additional features as it is included in the z800.
>>>>>
>>>>> Signed-off-by: Aurelien Jarno <address@hidden>
>>>>> ---
>>>>>  target/s390x/cpu_models.c | 13 ++++++-------
>>>>>  1 file changed, 6 insertions(+), 7 deletions(-)
>>>>>
>>>>> diff --git a/target/s390x/cpu_models.c b/target/s390x/cpu_models.c
>>>>> index fc3cb25cc3..c13bbd852c 100644
>>>>> --- a/target/s390x/cpu_models.c
>>>>> +++ b/target/s390x/cpu_models.c
>>>>> @@ -668,8 +668,6 @@ static void 
>>>>> add_qemu_cpu_model_features(S390FeatBitmap fbm)
>>>>>      static const int feats[] = {
>>>>>          S390_FEAT_STFLE,
>>>>>          S390_FEAT_EXTENDED_IMMEDIATE,
>>>>> -        S390_FEAT_EXTENDED_TRANSLATION_2,
>>>>> -        S390_FEAT_LONG_DISPLACEMENT,
>>>>>          S390_FEAT_LONG_DISPLACEMENT_FAST,
>>>>>          S390_FEAT_ETF2_ENH,
>>>>>          S390_FEAT_STORE_CLOCK_FAST,
>>>>> @@ -696,9 +694,9 @@ static S390CPUModel *get_max_cpu_model(Error **errp)
>>>>>      if (kvm_enabled()) {
>>>>>          kvm_s390_get_host_cpu_model(&max_model, errp);
>>>>>      } else {
>>>>> -        /* TCG emulates a z900 (with some optional additional features) 
>>>>> */
>>>>> -        max_model.def = &s390_cpu_defs[0];
>>>>> -        bitmap_copy(max_model.features, max_model.def->default_feat,
>>>>> +        /* TCG emulates a z800 (with some optional additional features) 
>>>>> */
>>>>> +        max_model.def = s390_find_cpu_def(0x2066, 7, 3, NULL);
>>>>> +        bitmap_copy(max_model.features, max_model.def->full_feat,
>>>>>                      S390_FEAT_MAX);
>>>
>>> This is most likely wrong: you're indicating features here that are not
>>> available on tcg. esp. S390_FEAT_SIE_F2 and friends.
>>>
>>> I think should only copy the base features and add whatever else is
>>> available via add_qemu_cpu_model_features() as already done.
>>
>> The patch series added all the z800 features exposed via STFL/STFLE.
>> Indeed the SIE features are missing, but anyway QEMU doesn't emulate SIE
>> at all so the lack of these features are not exposed to the guest. In that
>> regard QEMU already wrongly claim to emulate a z900.

Please note that:

a) SIE features were never part of the max model. QEMU never claimed
that. With your change one could suddenly do a -cpu z900,sie_f2=on,
which is wrong.

b) The SIE_F2 feature tells the guest that the SIE instruction is
available. E.g. Linux will look at this bit and show SIE support in
/proc/cpuinfo and unlock the KVM module.

Please, just don't add features to the MAX model that we don't implement.

>>
>>
>>>>>          add_qemu_cpu_model_features(max_model.features);
>>>>>      }
>>>>> @@ -956,8 +954,9 @@ static void s390_qemu_cpu_model_initfn(Object *obj)
>>>>>      S390CPU *cpu = S390_CPU(obj);
>>>>>  
>>>>>      cpu->model = g_malloc0(sizeof(*cpu->model));
>>>>> -    /* TCG emulates a z900 (with some optional additional features) */
>>>>> -    memcpy(&s390_qemu_cpu_defs, &s390_cpu_defs[0], 
>>>>> sizeof(s390_qemu_cpu_defs));
>>>>> +    /* TCG emulates a z800 (with some optional additional features) */
>>>>> +    memcpy(&s390_qemu_cpu_defs, s390_find_cpu_def(0x2066, 7, 3, NULL),
>>>>> +           sizeof(s390_qemu_cpu_defs));
>>>>
>>>> No changing the qemu model without compatibility handling.
>>>>
>>> Please have a look at the following mail for a possible solution:
>>>
>>> https://lists.gnu.org/archive/html/qemu-devel/2017-05/msg06030.html
>>>
>>> This could be moved to a separate patch. So this patch really should
>>> just care about the maximum model, not the qemu model.
>>
>> From what I understand from this thread, the patch from Thomas Huth was
>> finally considered acceptable. I am adding him in Cc: so that he can
>> comment.
> 
> In the 2nd version of my patch, I only changed the full_feat of the
> model definitions, but not the base_feat and default_feat fields, so
> that the CPU stays the same when you run QEMU without "-cpu" parameter
> (or simply with "-cpu qemu").
> 
> Consider that the following scenario should work: Start a QEMU v2.10
> with "-M s390-ccw-virtio-2.9" and a QEMU v2.9 with "-M
> s390-ccw-virtio-2.9 -incoming ...". Then it should be possible to
> migrate from the v2.10 to the v2.9 instance without problems. This won't
> work anymore if you changed the default feature bits unconditionally.
> 
>  Thomas
> 

The following can be used to make sure the model does not change:

s390x-softmmu/qemu-system-s390x -S -qmp stdio

-> { "execute": "qmp_capabilities" }
-> { "execute": "query-cpu-model-expansion", "arguments": { "type":
"static", "model": { "name": "qemu" } } }

With existing compat machines, this has to give
-> "{"return": {"model": {"name": "z900-base"}}}"


Also, is there any specific reason why to go for a z800 and not a z900.3
? (IBM zSeries 900 GA3).

z900 GA3 is the EC model, z800 GA1 is the BC model. Both have same
features and capabilities. EC always smells more powerful ;) In the end
it doesn't matter.

-- 

Thanks,

David