Re: [GIT PULL 1/1] KVM: s390: Remove false WARN_ON_ONCE for the PQAP instruction

[David Hildenbrand <david@redhat.com>]

On 06.05.20 15:04, Paolo Bonzini wrote:
> On 06/05/20 14:23, Christian Borntraeger wrote:
>>
>> the problem is that z/VM seems to have disabled that bit. The interpretion
>> only works when all layers have this bit set. (Its called an effective bit). 
>> So we have
>> LPAR -> ECA_APIE = 1
>> Z/VM -> ECA_APIE = 0
>> KVM  -> ECA_APIE = 1
>> nested KVM guest --> does during boot in ap_instructions_available PQAP with
>> FC==0 to test if crypto is available.
>>
>> As the nested guest is in fact a shadow guest of z/VM
>> this will now intercept to z/VM, which will forward that to KVM.
> 
> Right, and I'd understand, that since KVM has set ECA_APIE=1, z/VM
> should either:
> 
> * not forward the intercept to KVM unless FC==3
> 
> * toggle ECA_APIE back to 1 while running the KVM nested guest.
> 
> Of course I have no idea if either of these choices is impossible, or
> too expensive, but this is how we'd try to handle it for x86 features.
> 
> So it would be a z/VM bug, because it's not virtualizing ECA_APIE=1
> correctly.  The next question is, if removing the WARN_ON is okay for
> you, should KVM not bother setting ECA_APIE=1 so that you don't trigger
> the z/VM bug at all?
> 

IIRC, it's perfectly valid - according to the documentation - to ignore
ECA_APIE. That's the weird thing about effective controls. You don't
really know what you get in the end.

But, it's a long time since I had the chance to look at that
documentation ...

-- 
Thanks,

David / dhildenb