[PATCH] arm64: KVM: Support X-Gene guest VCPU on APM X-Gene host

[Alexander Graf]

On 15.07.2013, at 14:56, Anup Patel wrote:

> Hi Marc,
> 
> On Mon, Jul 15, 2013 at 6:09 PM, Marc Zyngier <marc.zyngier at arm.com> wrote:
>> Hi Anup,
>> 
>> On 15/07/13 12:46, Anup Patel wrote:
>>> This patch allows us to have X-Gene guest VCPU when using
>>> KVM arm64 on APM X-Gene host.
>>> 
>>> We add KVM_ARM_TARGET_XGENE_V8 for X-Gene compatible guest VCPU
>>> and we return KVM_ARM_TARGET_XGENE_V8 in kvm_target_cpu() when
>>> running on X-Gene host.
>>> 
>>> Signed-off-by: Anup Patel <anup.patel at linaro.org>
>>> Signed-off-by: Pranavkumar Sawargaonkar <pranavkumar at linaro.org>
>>> ---
>>> arch/arm64/include/uapi/asm/kvm.h    |    3 ++-
>>> arch/arm64/kvm/guest.c               |   34 ++++++++++++++++++++++------------
>>> arch/arm64/kvm/sys_regs_generic_v8.c |    3 +++
>>> 3 files changed, 27 insertions(+), 13 deletions(-)
>>> 
>>> diff --git a/arch/arm64/include/uapi/asm/kvm.h b/arch/arm64/include/uapi/asm/kvm.h
>>> index 5031f42..8194707 100644
>>> --- a/arch/arm64/include/uapi/asm/kvm.h
>>> +++ b/arch/arm64/include/uapi/asm/kvm.h
>>> @@ -55,8 +55,9 @@ struct kvm_regs {
>>> #define KVM_ARM_TARGET_AEM_V8                0
>>> #define KVM_ARM_TARGET_FOUNDATION_V8 1
>>> #define KVM_ARM_TARGET_CORTEX_A57    2
>>> +#define KVM_ARM_TARGET_XGENE_V8              3
>>> 
>>> -#define KVM_ARM_NUM_TARGETS          3
>>> +#define KVM_ARM_NUM_TARGETS          4
>>> 
>>> /* KVM_ARM_SET_DEVICE_ADDR ioctl id encoding */
>>> #define KVM_ARM_DEVICE_TYPE_SHIFT    0
>>> diff --git a/arch/arm64/kvm/guest.c b/arch/arm64/kvm/guest.c
>>> index 2c3ff67..e99b0a5 100644
>>> --- a/arch/arm64/kvm/guest.c
>>> +++ b/arch/arm64/kvm/guest.c
>>> @@ -207,19 +207,29 @@ int __attribute_const__ kvm_target_cpu(void)
>>>      unsigned long implementor = read_cpuid_implementor();
>>>      unsigned long part_number = read_cpuid_part_number();
>>> 
>>> -     if (implementor != ARM_CPU_IMP_ARM)
>>> -             return -EINVAL;
>>> -
>>> -     switch (part_number) {
>>> -     case ARM_CPU_PART_AEM_V8:
>>> -             return KVM_ARM_TARGET_AEM_V8;
>>> -     case ARM_CPU_PART_FOUNDATION:
>>> -             return KVM_ARM_TARGET_FOUNDATION_V8;
>>> -     case ARM_CPU_PART_CORTEX_A57:
>>> -             /* Currently handled by the generic backend */
>>> -             return KVM_ARM_TARGET_CORTEX_A57;
>>> +     switch (implementor) {
>>> +     case ARM_CPU_IMP_ARM:
>>> +             switch (part_number) {
>>> +             case ARM_CPU_PART_AEM_V8:
>>> +                     return KVM_ARM_TARGET_AEM_V8;
>>> +             case ARM_CPU_PART_FOUNDATION:
>>> +                     return KVM_ARM_TARGET_FOUNDATION_V8;
>>> +             case ARM_CPU_PART_CORTEX_A57:
>>> +                     return KVM_ARM_TARGET_CORTEX_A57;
>>> +             default:
>>> +                     return -EINVAL;
>>> +             }
>>> +             break;
>>> +     case ARM_CPU_IMP_APM:
>>> +             switch (part_number) {
>>> +             case APM_CPU_PART_POTENZA:
>>> +                     return KVM_ARM_TARGET_XGENE_V8;
>> 
>> Why don't we have KVM_ARM_TARGET_XGENE_POTENZA (or something similar)
>> instead? I don't expect all the X-Gene CPUs to be the same forever...
> 
> OK, I will rename it to KVM_ARM_TARGET_XGENE_POTENZA.
> 
> Does this mean that with every new ARM64 CPU we will have to add a new
> target for KVM ARM64 ?

Only for different core types, no? Any Cortex-A57 should still behave the same.

> If so then I think the list of targets will grow very fast.
> 
> I also realized that if we add a new target type in KVM ARM64 then we have
> to also update KVMTOOL to use the new target else KVMTOOL fails to
> recognize the target provided by KVM ARM64.

Right. It might make sense to have a fetch mechanism for the host cpu part. So you can ask KVM for the host cpu type and pass that back in here.

> Do you think we can have KVM_ARM_TARGET_xxx to represent a common
> target for a family of CPUs from given ARM64 vendor?

Anything that is compatible is compatible :). I don't know the product roadmaps for X-Gene cores, but you will want to make the field here as coarse grained as possible, while maintaining the guarantee that a guest still behaves the same.


Alex