Re: [PATCH v3 2/4] s390x: kvm: topology: interception of PTF instruction

[Thomas Huth <thuth@redhat.com>]

On 13/10/2021 09.55, Pierre Morel wrote:
> 
> 
> On 10/13/21 09:25, Thomas Huth wrote:
>> On 16/09/2021 15.50, Pierre Morel wrote:
>>> When the host supports the CPU topology facility, the PTF
>>> instruction with function code 2 is interpreted by the SIE,
>>> provided that the userland hypervizor activates the interpretation
>>> by using the KVM_CAP_S390_CPU_TOPOLOGY KVM extension.
>>>
>>> The PTF instructions with function code 0 and 1 are intercepted
>>> and must be emulated by the userland hypervizor.
>>>
>>> Signed-off-by: Pierre Morel <pmorel@linux.ibm.com>
>>> ---
...
>>> diff --git a/target/s390x/kvm/kvm.c b/target/s390x/kvm/kvm.c
>>> index 5b1fdb55c4..dd036961fe 100644
>>> --- a/target/s390x/kvm/kvm.c
>>> +++ b/target/s390x/kvm/kvm.c
>>> @@ -97,6 +97,7 @@
>>>   #define PRIV_B9_EQBS                    0x9c
>>>   #define PRIV_B9_CLP                     0xa0
>>> +#define PRIV_B9_PTF                     0xa2
>>>   #define PRIV_B9_PCISTG                  0xd0
>>>   #define PRIV_B9_PCILG                   0xd2
>>>   #define PRIV_B9_RPCIT                   0xd3
>>> @@ -362,6 +363,7 @@ int kvm_arch_init(MachineState *ms, KVMState *s)
>>>       kvm_vm_enable_cap(s, KVM_CAP_S390_USER_SIGP, 0);
>>>       kvm_vm_enable_cap(s, KVM_CAP_S390_VECTOR_REGISTERS, 0);
>>>       kvm_vm_enable_cap(s, KVM_CAP_S390_USER_STSI, 0);
>>> +    kvm_vm_enable_cap(s, KVM_CAP_S390_CPU_TOPOLOGY, 0);
>>
>> Should this maybe rather be done in the last patch, to avoid a state where 
>> PTF is available, but STSI 15 is not implemented yet (when bisecting 
>> through these commits later)?
>>
>>   Thomas
>>
> 
> Yes you are right, thanks.

I'm also still a little bit surprised that there is really no migration code 
involved here yet. What if a guest gets started on a system with 
KVM_CAP_S390_CPU_TOPOLOGY support and later migrated to a system without 
KVM_CAP_S390_CPU_TOPOLOGY support? Is there already some magic in place that 
rejects such a migration? If not, the guest might first learn that it could 
use the PTF instruction, but suddenly it is then not available anymore? Does 
Linux cope right with PTF becoming unavailable during runtime? But even if 
it does, I think it's likely not in the sense of the architecture if certain 
instructions might disappear during runtime? Or do I miss something?

  Thomas