Re: [Qemu-devel] [PATCH qemu] spapr/target-ppc/kvm: Only add hcall-instructions if KVM supports it

[Thomas Huth <thuth@redhat.com>]

On 16.03.2016 03:43, Alexey Kardashevskiy wrote:
> On 03/15/2016 10:32 PM, Thomas Huth wrote:
>> On 15.03.2016 10:42, Alexey Kardashevskiy wrote:
>>> On 03/15/2016 07:18 PM, Thomas Huth wrote:
>>>>
>>>>    Hi Alexey,
>>>>
>>>> On 15.03.2016 06:51, Alexey Kardashevskiy wrote:
>>>>> ePAPR defines "hcall-instructions" device-tree property which contains
>>>>> code to call hypercalls in ePAPR paravirtualized guests. However this
>>>>> property is also present for pseries guests where it does not make
>>>>> sense,
>>>>> even though it contains dummy code which simply fails.
>>>>>
>>>>> Instead of maintaining the property (which used to be BE only; then
>>>>> was
>>>>> fixed to be endian-agnostic) and confusing the guest (which might
>>>>> think
>>>>> there is ePAPR host while there is none), this simply does not
>>>>> the property to the device tree if the host kernel does not implement
>>>>> it.
>>>>>
>>>>> In order to tell the machine code if the host kernel supports
>>>>> KVM_CAP_PPC_GET_PVINFO, this changes kvmppc_get_hypercall() to
>>>>> return 1
>>>>> if the host kernel does not implement it (which is HV KVM case).
>>>>>
>>>>> Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
>>>>> ---
>>>>>
>>>>>
>>>>> Alexander,
>>>>>
>>>>> We just got a bug report that LE guests would not boot under quite
>>>>> old QEMU
>>>>> and we (powerkvm) wonder if it makes sense to backport endian-agnostic
>>>>> hypercall code to older QEMU or it is simpler/more correct
>>>>> not to have epapr-hypercall property in the tree.
>>>>>
>>>>>
>>>>> ---
>>>>>    hw/ppc/spapr.c   | 9 +++++----
>>>>>    target-ppc/kvm.c | 2 +-
>>>>>    2 files changed, 6 insertions(+), 5 deletions(-)
>>>>>
>>>>> diff --git a/hw/ppc/spapr.c b/hw/ppc/spapr.c
>>>>> index 43708a2..8130eb4 100644
>>>>> --- a/hw/ppc/spapr.c
>>>>> +++ b/hw/ppc/spapr.c
>>>>> @@ -497,10 +497,11 @@ static void *spapr_create_fdt_skel(hwaddr
>>>>> initrd_base,
>>>>>                 * Older KVM versions with older guest kernels were
>>>>> broken with the
>>>>>                 * magic page, don't allow the guest to map it.
>>>>>                 */
>>>>> -            kvmppc_get_hypercall(first_cpu->env_ptr, hypercall,
>>>>> -                                 sizeof(hypercall));
>>>>> -            _FDT((fdt_property(fdt, "hcall-instructions", hypercall,
>>>>> -                              sizeof(hypercall))));
>>>>> +            if (!kvmppc_get_hypercall(first_cpu->env_ptr, hypercall,
>>>>> +                                      sizeof(hypercall))) {
>>>>> +                _FDT((fdt_property(fdt, "hcall-instructions",
>>>>> hypercall,
>>>>> +                                   sizeof(hypercall))));
>>>>> +            }
>>>>>            }
>>>>>            _FDT((fdt_end_node(fdt)));
>>>>>        }
>>>>> diff --git a/target-ppc/kvm.c b/target-ppc/kvm.c
>>>>> index 776336b..e5183db 100644
>>>>> --- a/target-ppc/kvm.c
>>>>> +++ b/target-ppc/kvm.c
>>>>> @@ -2001,7 +2001,7 @@ int kvmppc_get_hypercall(CPUPPCState *env,
>>>>> uint8_t *buf, int buf_len)
>>>>>        hc[2] = cpu_to_be32(0x48000008);
>>>>>        hc[3] = cpu_to_be32(bswap32(0x3860ffff));
>>>>>
>>>>> -    return 0;
>>>>> +    return 1;
>>>>>    }
>>>>>
>>>>>    static inline int kvmppc_enable_hcall(KVMState *s, target_ulong
>>>>> hcall)
>>>>
>>>> Sorry, I have a hard time to understand what this is really good
>>>> for. Is
>>>> it a patch for current QEMU or for older ones? If it is for older ones,
>>>> then why did you not CC: to qemu-stable?
>>>> If it is for current QEMU, then I've got some more questions about
>>>> things I do not understand:
>>>>
>>>> 1) In your patch description, you talk about ePAPR and that the
>>>> property
>>>> does not make sense for pseries. But why is this code then available at
>>>> all in spapr.c? ... there must be a reason for this, I think (like
>>>> using
>>>> a different h-call on nested KVM-PR for example?)
>>>
>>>
>>> No, this is from old times when there was only PR KVM fully emulating
>>> powermac (not pseries) which needed to interact with the hypervisor and
>>> epapr_hypercall was chosen for this.
>>>
>>>
>>>> 2) The code in spapr.c is already protected with a
>>>>     if (kvmppc_has_cap_fixup_hcalls()) ...
>>>> and that CAP should only be there if the PVINFO CAP is available, too.
>>>> So I don't see how you could run into that problem anyway where PVINFO
>>>> is _not_ available but the FIXUP_HCALL CAP _is_ available?
>>>
>>>
>>> HV KVM guest calls (on pseries machine as well):
>>>
>>> kvm_guest_init
>>> kvm_para_has_feature
>>> kvm_arch_para_features
>>> kvm_para_available - this returns "1"
>>> epapr_hypercall0_1(KVM_HC_FEATURES)
>>>
>>> This epapr_hypercall0_1() calls a binary blob from "hcall-instructions".
>>> And fails if the guest is LE and the blob from BE-only times.
>>
>> What about that "if (kvmppc_has_cap_fixup_hcalls())" ? Could you please
>> check why this succeeds on your system , but the KVM_CAP_PPC_GET_PVINFO
>> call does not?
> 
> KVM_CAP_PPC_FIXUP_HCALL is always enabled for CONFIG_PPC_BOOK3S_64,
> KVM_CAP_PPC_GET_PVINFO is only enabled for "!hv_enabled".

Ah, that's the detail that I missed. Thanks a lot for the hint!
... ok, then I think your patch is the right thing to do, but you should
maybe change the patch description a little bit (since this call still
might make sense on sPAPR, too).

 Thomas