[PATCH v7 19/19] KVM: ARM64: Add a new kvm ARM PMU device

Marc Zyngier marc.zyngier at arm.com
Thu Dec 17 02:38:09 PST 2015


On 17/12/15 10:10, Shannon Zhao wrote:
> 
> 
> On 2015/12/17 17:38, Marc Zyngier wrote:
>> On 17/12/15 08:41, Shannon Zhao wrote:
>>>>
>>>>
>>>> On 2015/12/17 16:33, Marc Zyngier wrote:
>>>>>> On Thu, 17 Dec 2015 15:22:50 +0800
>>>>>> Shannon Zhao <zhaoshenglong at huawei.com> wrote:
>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> On 2015/12/17 4:33, Christoffer Dall wrote:
>>>>>>>>>>>>>> On Wed, Dec 16, 2015 at 04:06:49PM +0800, Shannon Zhao wrote:
>>>>>>>>>>>>>>>>>> Hi,
>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>> On 2015/12/16 15:31, Shannon Zhao wrote:
>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> But in this case, you're returning an error if it is *not* initialized.
>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> I understand that in that case you cannot return an interrupt number (-1
>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> would be weird), but returning -EBUSY feels even more weird.
>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> I'd settle for -ENOXIO, or something similar. Anyone having a better idea?
>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> ENXIO or ENODEV would be my choice too, and add that to the
>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> Documentation clearly describing when this error code is used.
>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> By the way, why do you loop over all VCPUS to set the same value when
>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> you can't do anything per VCPU anyway?  It seems to me it's either a
>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> per-VM property (that you can store on the VM data structure) or it's a
>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> true per-VCPU property?
>>>>>>>>>>>>>>>>>>>>>> This is a per-VCPU property. PMU interrupt could be PPI or SPI. For PPI
>>>>>>>>>>>>>>>>>>>>>> the interrupt numbers are same for each vcpu, while for SPI they are
>>>>>>>>>>>>>>>>>>>>>> different, so it needs to set them separately. I planned to support both
>>>>>>>>>>>>>>>>>>>>>> PPI and SPI. I think I should add support for SPI at this moment and let
>>>>>>>>>>>>>>>>>>>>>> users (QEMU) to set these interrupts for each one.
>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>> How about below vPMU Documentation?
>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>> ARM Virtual Performance Monitor Unit (vPMU)
>>>>>>>>>>>>>>>>>> ===========================================
>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>> Device types supported:
>>>>>>>>>>>>>>>>>>   KVM_DEV_TYPE_ARM_PMU_V3         ARM Performance Monitor Unit v3
>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>> Instantiate one PMU instance for per VCPU through this API.
>>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>> Groups:
>>>>>>>>>>>>>>>>>>   KVM_DEV_ARM_PMU_GRP_IRQ
>>>>>>>>>>>>>>>>>>   Attributes:
>>>>>>>>>>>>>>>>>>     The attr field of kvm_device_attr encodes two values:
>>>>>>>>>>>>>>>>>>     bits:     | 63 .... 32 | 31 .... 0 |
>>>>>>>>>>>>>>>>>>     values:   | vcpu_index |  irq_num  |
>>>>>>>>>> BTW, I change this Attribute to below format and pass vcpu_index through
>>>>>>>>>> this Attribute while pass irq_num through kvm_device_attr.addr.
>>>>>>>>>> Is it fine?
>>>>>>>>>>
>>>>>>>>>>     bits:     | 63 .... 32 | 31 ....  0 |
>>>>>>>>>>     values:   |  reserved  | vcpu_index |
>>>>>> Using the .addr field for something that is clearly not an address is
>>>>>> rather odd. Is there any prior usage of that field for something that
>>>>>> is not an address?
>>>>
>>>> I see this usage in vgic_attr_regs_access(). But if you prefer previous
>>>> one, I'll use that.
>> Ah, you're using the .addr field to point to a userspace value, not to
>> carry the value itself! That'd be fine by me (even if I still prefer the
>> original one), but I'd like others to chime in (I'm quite bad at
>> defining userspace stuff...).
> 
> Another reason I think is that it needs to pass the irq_num to user
> space when calling get_attr. It could be passed via kvm_device_attr.addr
> while can't be passed via kvm_device_attr.attr within current framework.

That's a very good reason. Go for it.

Thanks,

	M.
-- 
Jazz is not dead. It just smells funny...



More information about the linux-arm-kernel mailing list