Re: [PATCH 7/9] KVM: Expose the architectural performance monitoring CPUID leaf

[Gleb Natapov <gleb@redhat.com>]

On Tue, Nov 01, 2011 at 10:24:41AM -0600, David Ahern wrote:
> 
> 
> On 11/01/2011 10:18 AM, Gleb Natapov wrote:
> > On Tue, Nov 01, 2011 at 09:49:27AM -0600, David Ahern wrote:
> >> On 10/30/2011 10:53 AM, Gleb Natapov wrote:
> >>> From: Avi Kivity <avi@redhat.com>
> >>>
> >>> Provide a CPUID leaf that describes the emulated PMU.
> >>>
> >>> Signed-off-by: Avi Kivity <avi@redhat.com>
> >>> Signed-off-by: Gleb Natapov <gleb@redhat.com>
> >>> ---
> >>>  arch/x86/kvm/x86.c |   23 ++++++++++++++++++++++-
> >>>  1 files changed, 22 insertions(+), 1 deletions(-)
> >>>
> >>> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> >>> index 5ea4cb8..56153a9 100644
> >>> --- a/arch/x86/kvm/x86.c
> >>> +++ b/arch/x86/kvm/x86.c
> >>> @@ -2543,6 +2543,28 @@ static void do_cpuid_ent(struct kvm_cpuid_entry2 *entry, u32 function,
> >>>  	}
> >>>  	case 9:
> >>>  		break;
> >>> +	case 0xa: { /* Architectural Performance Monitoring */
> >>> +		struct x86_pmu_capability cap;
> >>> +
> >>> +		perf_get_x86_pmu_capability(&cap);
> >>> +
> >>> +		/*
> >>> +		 * Only support guest architectural pmu on a host
> >>> +		 * with architectural pmu.
> >>> +		 */
> >>> +		if (!cap.version)
> >>> +			memset(&cap, 0, sizeof(cap));
> >>> +
> >>> +		entry->eax = min(cap.version, 2)
> >>> +			| (cap.num_counters_gp << 8)
> >>> +			| (cap.bit_width_gp << 16)
> >>> +			| (cap.events_mask_len << 24);
> >>> +		entry->ebx = cap.events_mask;
> >>> +		entry->ecx = 0;
> >>> +		entry->edx = cap.num_counters_fixed
> >>> +			| (cap.bit_width_fixed << 5);
> >>> +		break;
> >>> +	}
> >>
> >> If PERF_EVENTS is disabled in the host kernel will KVM return the right
> >> thing for the guest that might have PERF_EVENTS enabled?
> >>
> > With correct implementation of perf_get_x86_pmu_capability() it should.
> > PERF_EVENTS disable version should set cap.version to zero, so guest will
> > find that vcpu does not provide architectural PMU.
> 
> Ok.
> 
> I was also thinking about kvm-kmod where newer KVM source is compiled
> against older kernels - which at one point had PERF_EVENTS as a
> configurable parameter. In this case it is possible to have perf events
> disabled host side yet enabled guest side.
> 
KVM PMU implementation uses host perf events to talk to HW. Without host
perf infrastructure KVM will not be able to provide PMU to a guest. Or
are you saying that we should support compiling KVM with !PERF_EVENTS? I'd
rather just make KVM select PERF_EVENTS.

--
			Gleb.