Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Tue, 19 Jan 2021 17:13:47 -0800
From:   Sean Christopherson <seanjc@google.com>
To:     "Xu, Like" <like.xu@intel.com>
Cc:     Paolo Bonzini <pbonzini@redhat.com>,
        Stephane Eranian <eranian@google.com>,
        Vitaly Kuznetsov <vkuznets@redhat.com>,
        Wanpeng Li <wanpengli@tencent.com>,
        Jim Mattson <jmattson@google.com>,
        Joerg Roedel <joro@8bytes.org>, kvm@vger.kernel.org,
        linux-kernel@vger.kernel.org, Like Xu <like.xu@linux.intel.com>
Subject: Re: [PATCH] KVM: x86/pmu: Fix HW_REF_CPU_CYCLES event
 pseudo-encoding in intel_arch_events[]
Message-ID: <YAeDy66f13uLMrwU@google.com>
References: <20201230081916.63417-1-like.xu@linux.intel.com>
 <1ff5381c-3057-7ca2-6f62-bbdcefd8e427@linux.intel.com>
 <YAHRMK5SmrmMx8hg@google.com>
 <b3623ea4-b2a1-e825-68f9-d97a6e7a07f4@intel.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <b3623ea4-b2a1-e825-68f9-d97a6e7a07f4@intel.com>
Precedence: bulk

On Mon, Jan 18, 2021, Xu, Like wrote:
> On 2021/1/16 1:30, Sean Christopherson wrote:
> > On Fri, Jan 15, 2021, Like Xu wrote:
> > > > diff --git a/arch/x86/kvm/vmx/pmu_intel.c b/arch/x86/kvm/vmx/pmu_intel.c
> > > > index a886a47daebd..013e8d253dfa 100644
> > > > --- a/arch/x86/kvm/vmx/pmu_intel.c
> > > > +++ b/arch/x86/kvm/vmx/pmu_intel.c
> > > > @@ -29,7 +29,7 @@ static struct kvm_event_hw_type_mapping intel_arch_events[] = {
> > > >    	[4] = { 0x2e, 0x41, PERF_COUNT_HW_CACHE_MISSES },
> > > >    	[5] = { 0xc4, 0x00, PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
> > > >    	[6] = { 0xc5, 0x00, PERF_COUNT_HW_BRANCH_MISSES },
> > > > -	[7] = { 0x00, 0x30, PERF_COUNT_HW_REF_CPU_CYCLES },
> > > > +	[7] = { 0x00, 0x03, PERF_COUNT_HW_REF_CPU_CYCLES },
> > In a follow up patch, would it be sane/appropriate to define these magic numbers
> > in asm/perf_event.h and share them between intel_perfmon_event_map and
> > intel_arch_events?  Without this patch, it's not at all obvious that these are
> > intended to align with the Core (arch?) event definitions.
> 
> The asm/perf_event.h is x86 generic and svm has a amd_perfmon_event_map.

Ugh, duh.

> How about adding an interface similar to perf_get_x86_pmu_capability()
> so that we can use magic numbers directly from the host perf ?
> (it looks we may have a performance drop, compared to static array)

Alternatively, you could use the existing event_map() to generate
intel_arch_events[] during init.  That might be easier since, AFAICT, the array
indices have different meaning for KVM than for perf.

Honestly, unless there are going to be new arch events in the near-ish future,
it may not be worth the effort at this point.  Until now, the above table hadn't
changed in over five years.  I.e. don't put a bunch of effort into this unless
you want to :-)