答复: [PATCH] [RFC] kvm: x86: emulate APERF/MPERF registers

["Li,Rongqing" <lirongqing@baidu.com>]

> -----邮件原件-----
> 发件人: Sean Christopherson [mailto:sean.j.christopherson@intel.com]
> 发送时间: 2020年4月24日 22:46
> 收件人: Peter Zijlstra <peterz@infradead.org>
> 抄送: Li,Rongqing <lirongqing@baidu.com>; linux-kernel@vger.kernel.org;
> kvm@vger.kernel.org; x86@kernel.org; hpa@zytor.com; bp@alien8.de;
> mingo@redhat.com; tglx@linutronix.de; joro@8bytes.org;
> jmattson@google.com; wanpengli@tencent.com; vkuznets@redhat.com;
> pbonzini@redhat.com
> 主题: Re: [PATCH] [RFC] kvm: x86: emulate APERF/MPERF registers
> 
> On Fri, Apr 24, 2020 at 12:01:43PM +0200, Peter Zijlstra wrote:
> > On Fri, Apr 24, 2020 at 01:08:55PM +0800, Li RongQing wrote:
> > > diff --git a/arch/x86/kvm/cpuid.c b/arch/x86/kvm/cpuid.c index
> > > 901cd1fdecd9..00e4993cb338 100644
> > > --- a/arch/x86/kvm/cpuid.c
> > > +++ b/arch/x86/kvm/cpuid.c
> > > @@ -558,7 +558,10 @@ static inline int __do_cpuid_func(struct
> kvm_cpuid_array *array, u32 function)
> > >  	case 6: /* Thermal management */
> > >  		entry->eax = 0x4; /* allow ARAT */
> > >  		entry->ebx = 0;
> > > -		entry->ecx = 0;
> > > +		if (boot_cpu_has(X86_FEATURE_APERFMPERF))
> > > +			entry->ecx = 0x1;
> > > +		else
> > > +			entry->ecx = 0x0;
> > >  		entry->edx = 0;
> > >  		break;
> > >  	/* function 7 has additional index. */
> >
> > AFAICT this is generic x86 code, that is, this will tell an AMD SVM
> > guest it has APERFMPERF on.
> >
> > > diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c index
> > > 91749f1254e8..f20216fc0b57 100644
> > > --- a/arch/x86/kvm/vmx/vmx.c
> > > +++ b/arch/x86/kvm/vmx/vmx.c
> > > @@ -1064,6 +1064,11 @@ static inline void pt_save_msr(struct pt_ctx
> > > *ctx, u32 addr_range)
> > >
> > >  static void pt_guest_enter(struct vcpu_vmx *vmx)  {
> > > +	struct kvm_vcpu *vcpu = &vmx->vcpu;
> > > +
> > > +	rdmsrl(MSR_IA32_MPERF, vcpu->arch.host_mperf);
> > > +	rdmsrl(MSR_IA32_APERF, vcpu->arch.host_aperf);
> 
> Why are these buried in Processor Trace code?  Is APERFMERF in any way
> dependent on PT?
> 

No;
I will move it out PT codes

> > > +
> > >  	if (vmx_pt_mode_is_system())
> > >  		return;
> > >
> > > @@ -1081,6 +1086,15 @@ static void pt_guest_enter(struct vcpu_vmx
> > > *vmx)
> > >
> > >  static void pt_guest_exit(struct vcpu_vmx *vmx)  {
> > > +	struct kvm_vcpu *vcpu = &vmx->vcpu;
> > > +	u64 perf;
> > > +
> > > +	rdmsrl(MSR_IA32_MPERF, perf);
> > > +	vcpu->arch.v_mperf += perf - vcpu->arch.host_mperf;
> > > +
> > > +	rdmsrl(MSR_IA32_APERF, perf);
> > > +	vcpu->arch.v_aperf += perf - vcpu->arch.host_aperf;
> 
> This requires four RDMSRs per VMX transition.  Doing that unconditionally will
> drastically impact performance.  Not to mention that reading the MSRs
> without checking for host support will generate #GPs and WARNs on hardware
> without APERFMPERF.
> 
> Assuming we're going forward with this, at an absolute minimum the RDMSRs
> need to be wrapped with checks on host _and_ guest support for the emulated
> behavior.  Given the significant overhead, this might even be something that
> should require an extra opt-in from userspace to enable.
> 

will do as your suggestion 

> > > +
> > >  	if (vmx_pt_mode_is_system())
> > >  		return;
> > >
> > > @@ -1914,6 +1928,12 @@ static int vmx_get_msr(struct kvm_vcpu *vcpu,
> struct msr_data *msr_info)
> > >  		    !guest_cpuid_has(vcpu, X86_FEATURE_RDTSCP))
> > >  			return 1;
> > >  		goto find_shared_msr;
> > > +	case MSR_IA32_MPERF:
> > > +		msr_info->data = vcpu->arch.v_mperf;
> > > +		break;
> > > +	case MSR_IA32_APERF:
> > > +		msr_info->data = vcpu->arch.v_aperf;
> > > +		break;
> > >  	default:
> > >  	find_shared_msr:
> > >  		msr = find_msr_entry(vmx, msr_info->index);
> >
> > But then here you only emulate it for VMX, which then results in SVM
> > guests going wobbly.
> 
> Ya.

I will make this code suitable for AMD and Intel cpu

> 
> > Also, on Intel, the moment you advertise APERFMPERF, we'll try and
> > read MSR_PLATFORM_INFO / MSR_TURBO_RATIO_LIMIT*, I don't suppose
> > you're passing those through as well?
> 
> AFAICT, the proposed patch isn't fully advertising APERFMPERF, it's advertising
> Turbo Boost / Dynamic Acceleration to the guest when APERFMPERF can be
> used by the host to emulate IDA.  The transliteration of the above code to be
> VMX-only is:

Do you means when we forward APERFMPERF to guest , guest has similar IDA capability,
 and not need to consider MSR_PLATFORM_INFO / MSR_TURBO_RATIO_LIMIT* ?

-Li
> 
> diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c index
> 766303b31949..7e459b66b06e 100644
> --- a/arch/x86/kvm/vmx/vmx.c
> +++ b/arch/x86/kvm/vmx/vmx.c
> @@ -7191,6 +7191,9 @@ static __init void vmx_set_cpu_caps(void)
>         if (nested)
>                 kvm_cpu_cap_set(X86_FEATURE_VMX);
> 
> +       if (boot_cpu_has(X86_FEATURE_APERFMPERF))
> +               kvm_cpu_cap_set(X86_FEATURE_IDA);
> +
>         /* CPUID 0x7 */
>         if (kvm_mpx_supported())
>                 kvm_cpu_cap_check_and_set(X86_FEATURE_MPX);
> 
> I have no clue as to whether that's sane/correct.