Re: [PATCH v3 1/3] s390x: KVM: accept STSI for CPU topology information

[Pierre Morel <pmorel@linux.ibm.com>]

On 9/6/21 8:14 PM, David Hildenbrand wrote:
> On 01.09.21 11:43, Pierre Morel wrote:
>>
>>
>> On 8/31/21 3:59 PM, David Hildenbrand wrote:
>>> On 03.08.21 10:26, Pierre Morel wrote:
>>>> STSI(15.1.x) gives information on the CPU configuration topology.
>>>> Let's accept the interception of STSI with the function code 15 and
>>>> let the userland part of the hypervisor handle it when userland
>>>> support the CPU Topology facility.
>>>>
>>>> Signed-off-by: Pierre Morel <pmorel@linux.ibm.com>
>>>> ---
>>>>    arch/s390/kvm/priv.c | 7 ++++++-
>>>>    1 file changed, 6 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/arch/s390/kvm/priv.c b/arch/s390/kvm/priv.c
>>>> index 9928f785c677..8581b6881212 100644
>>>> --- a/arch/s390/kvm/priv.c
>>>> +++ b/arch/s390/kvm/priv.c
>>>> @@ -856,7 +856,8 @@ static int handle_stsi(struct kvm_vcpu *vcpu)
>>>>        if (vcpu->arch.sie_block->gpsw.mask & PSW_MASK_PSTATE)
>>>>            return kvm_s390_inject_program_int(vcpu, PGM_PRIVILEGED_OP);
>>>> -    if (fc > 3) {
>>>> +    if ((fc > 3 && fc != 15) ||
>>>> +        (fc == 15 && !test_kvm_facility(vcpu->kvm, 11))) {
>>>>            kvm_s390_set_psw_cc(vcpu, 3);
>>>>            return 0;
>>>>        }
>>>> @@ -893,6 +894,10 @@ static int handle_stsi(struct kvm_vcpu *vcpu)
>>>>                goto out_no_data;
>>>>            handle_stsi_3_2_2(vcpu, (void *) mem);
>>>>            break;
>>>> +    case 15:
>>>> +        trace_kvm_s390_handle_stsi(vcpu, fc, sel1, sel2, operand2);
>>>> +        insert_stsi_usr_data(vcpu, operand2, ar, fc, sel1, sel2);
>>>> +        return -EREMOTE;
>>>>        }
>>>>        if (kvm_s390_pv_cpu_is_protected(vcpu)) {
>>>>            memcpy((void *)sida_origin(vcpu->arch.sie_block), (void 
>>>> *)mem,
>>>>
>>>
>>> Sorry, I'm a bit rusty on s390x kvm facility handling.
>>>
>>>
>>> For test_kvm_facility() to succeed, the facility has to be in both:
>>>
>>> a) fac_mask: actually available on the HW and supported by KVM
>>> (kvm_s390_fac_base via FACILITIES_KVM, kvm_s390_fac_ext via
>>> FACILITIES_KVM_CPUMODEL)
>>>
>>> b) fac_list: enabled for a VM
>>>
>>> AFAIU, facility 11 is neither in FACILITIES_KVM nor
>>> FACILITIES_KVM_CPUMODEL, and I remember it's a hypervisor-managed bit.
>>>
>>> So unless we unlock facility 11 in FACILITIES_KVM_CPUMODEL, will
>>> test_kvm_facility(vcpu->kvm, 11) ever successfully trigger here?
>>>
>>>
>>> I'm pretty sure I am messing something up :)
>>>
>>
>> I think it is the same remark that Christian did as wanted me to use the
>> arch/s390/tools/gen_facilities.c to activate the facility.
>>
>> The point is that CONFIGURATION_TOPOLOGY, STFL, 11, is already defined
>> inside QEMU since full_GEN10_GA1, so the test_kvm_facility() will
>> succeed with the next patch setting the facility 11 in the mask when
>> getting the KVM_CAP_S390_CPU_TOPOLOGY from userland.
> 
> Ok, I see ...
> 
> QEMU knows the facility and as soon as we present it to QEMU, QEMU will 
> want to automatically enable it in the "host" model.
> 
> However, we'd like QEMU to join in and handle some part of it.
> 
> So indeed, handling it like KVM_CAP_S390_VECTOR_REGISTERS or 
> KVM_CAP_S390_RI looks like a reasonable approach.
> 
>>
>> But if we activate it in KVM via any of the FACILITIES_KVM_xxx in the
>> gen_facilities.c we will activate it for the guest what ever userland
>> hypervizor we have, including old QEMU which will generate an exception.
>>
>>
>> In this circumstances we have the choice between:
>>
>> - use FACILITY_KVM and handle everything in kernel
>> - use FACILITY_KVM and use an extra CAPABILITY to handle part in kernel
>> to avoid guest crash and part in userland
> 
> This sounds quite nice to me. Implement minimal kernel support and 
> indicate the facility via stfl to user space.
> 
> In addition, add a new capability that intercepts to user space instead.
> 
> 
> ... but I can understand that it might not be worth it.

yes, since we need a CAPABILITY anyway I find it makes things more 
complicated.
> 
> 
> This patch as it stands doesn't make any sense on its own. Either 
> document how it's supposed to work and why it is currently dead code, or 
> simply squash into the next patch (preferred IMHO).
> 

Yes, you are right, I will squash it with the next patch.

Thanks,
Pierre

-- 
Pierre Morel
IBM Lab Boeblingen