Re: [Qemu-devel] [Qemu-ppc] KVM-PR is broken with current QEMU

[Cédric Le Goater]

On 09/22/2016 09:18 AM, Thomas Huth wrote:
> On Thu, 22 Sep 2016 07:30:52 +0200
> Thomas Huth <address@hidden> wrote:
> 
>> On Thu, 22 Sep 2016 11:57:15 +1000
>> David Gibson <address@hidden> wrote:
>>
>>> On Wed, Sep 21, 2016 at 10:22:11AM +0200, Thomas Huth wrote:
>>>> On 20.09.2016 16:39, Cédric Le Goater wrote:
>>>>> On 09/20/2016 04:24 PM, Thomas Huth wrote:
>>>>>> On 20.09.2016 16:04, Cédric Le Goater wrote:
>>>> [...]
>>>>>>> There are other issues after in the guest (kernel crashing). But I think
>>>>>>> these are related to TM which is not supported in KVM-PR. I am not sure
>>>>>>> where we are on that point.
>>>>>>
>>>>>> There was a patch some months ago:
>>>>>>
>>>>>> https://lists.gnu.org/archive/html/qemu-ppc/2016-04/msg00046.html
>>>>>>
>>>>>> ... but I think it has never been included, as far as I can see.
>>>>>
>>>>> and with that patch, the guest fully boots. But David had some concerns
>>>>> on the way it is done. It would be nice to put some cycle on this. 
>>>>
>>>> Looking at the mail thread, I think TM should be currently disabled for
>>>> both, KVM-PR and TCG, i.e. only enabled for KVM-HV. The TM support in
>>>> TCG is just fake, since TBEGIN always fails.
>>>
>>> Right.  So there's two questions here
>>>
>>> 1) Is qemu correctly advertising availability of TM in the device
>>> tree?
>>
>> If I've got that right, it's currently always advertising TM, even if
>> it's not really available (in TCG mode and PR mode).
>>
>>> If not we need to fix that, which might involve adding a kernel
>>> capability for the PR case.
>>>
>>> 2) Is the kvm unit test properly checking for availability of TM
>>> before executing?
>>
>> Not yet. That's why it would be good to get a proper way for testing
>> for the availability of TM --> i.e. something like Anton's patch.
>>
>>>> Once we've got proper TM support in TCG, this can be easily changed
>>>> within QEMU. And once we've got TM support in KVM-PR, I think we should
>>>> also introduce a capability flag to KVM which can be used to inform QEMU
>>>> about this.
>>>>
>>>> So I think Anton's patch currently just lacks the check for TCG.
>>>> Anton, if you've got some spare minutes, could you maybe send an updated
>>>> version of that patch?
>>>
>>> Sorry, which patch of Anton's?
>>
>> This one:
>> https://lists.gnu.org/archive/html/qemu-ppc/2016-04/msg00415.html
> 
> Actually, looking at that whole pa-feature code in QEMU, I think
> there's some more work to do here: Everything that is not using
> mmu_model==POWERPC_MMU_2_06 is automatically getting pa_features_207.
> This is sometimes completely wrong, for example when running with
> KVM-PR, the mmu_model for POWER7 is POWERPC_MMU_2_06a instead.
> Or when running with TCG, I think it's also perfectly legal to run the
> pseries machine with a POWER5+ or PPC970 CPU - and we certainly do not
> want to use pa_features_207 there.
> 
> So if you like, I can try to come up with a small patch series that
> cleans up this mess - and I could also include an updated versions of
> Anton's patch there unless he wants to redo the changes on his own...?
> 
>  Thomas
> 

That would be nice. I just gave a quick try on a f24/le kvm-pr running 
under a f24/le kvm-hv running under powernv. Only your couple of patches 
plus Anton's are needed to make it work.


Thanks,

C.