Re: [PATCH V4 2/4] KVM: x86/vPMU: Create vPMU interface for VMX and SVM

[Wei Huang <wei@redhat.com>]

On 06/10/2015 01:05 PM, Joerg Roedel wrote:
> On Wed, Jun 10, 2015 at 11:43:20AM -0500, Wei Huang wrote:
>> On 06/10/2015 05:12 AM, Joerg Roedel wrote:
>>> On Fri, Jun 05, 2015 at 01:20:14AM -0400, Wei Huang wrote:
>>>> +
>>>> +	struct kvm_pmu_ops *(*get_pmu_ops)(void);
>>>
>>> Can't you just set kvm_pmu_ops in svm.c and vmx.c and save this
>>> call-back? Besides that the patch looks good.
>> Hi Joerg,
>>
>> Thanks for your review. How about setting up kvm_pmu_ops in
>> .hardware_setup function of VMX and SVM? More specifically:
>>
>> 1) EXPORT_SYMBOL_GPL(kvm_pmu_ops) from pmu.c file;
>> 2) In vmx.c, set "kvm_pmu_ops = &intel_pmu_ops" in hardware_setup();
>> 3) In svm.c, set "kvm_pmu_ops = &amd_pmu_ops" in svm_hardware_setup().
>>
>> With that, we can get rid of the call-back.
> 
> Okay, just had a look at how this works with kvm_x86_ops, because my
> suggestion needs the EXPORT_SYMBOL_GPL(kvm_pmu_ops), which is better
> also avoided.
> 
> So how about putting the pmu-ops directly into the kvm_x86_ops as a
> (const) member, 
I like this idea better. Here is the (expanded) design:

1) add const "struct kvm_pmu_ops *pmu_ops" to kvm_x86_ops;
2) In VMX, "vmx_x86_ops.pmu_ops = &intel_pmu_ops";
3) In SVM, "svm_x86_ops.pmu_ops = &amd_pmu_ops".

Common kvm_pmu_xxxx functions (in pmu.c) will change accordingly
afterwards. If no objection, I will do this way.

or alternativly as an additional parameter to kvm_init?
> This still saves us the access functions.

Passing pmu-ops to kvm_init() implies that we have to change the public
API (adding a new parameter) to kvm_init, which will affect many
architectures. I will try to avoid it unless have to.

> 
> 
> 	Joerg
>