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;
IronPort-SDR: eywTvYHzDjkPog1u8P4uJqRr5jKt82OHoCEjbA8JMknfgCMRDRpIrv6ptyC34Pp/2XHpGRXzmf
 Cs1CcbFwUjGg==
IronPort-SDR: lvEpJOC50Hm1QMcKm0qilThYshi+I5TJTmGGWAwOLMchzbZnR3x0Ld8AXoUhQhceEgHmn6XE/y
 ODRLedmDsamA==
Subject: Re: [PATCH v2] KVM: x86/pmu: Reduce counter period change overhead
 and delay the effective time
From:   Like Xu <like.xu@linux.intel.com>
To:     pbonzini@redhat.com
Cc:     ehankland@google.com, jmattson@google.com, kvm@vger.kernel.org,
        linux-kernel@vger.kernel.org, wanpengli@tencent.com
References: <20200317075315.70933-1-like.xu@linux.intel.com>
 <20200317081458.88714-1-like.xu@linux.intel.com>
 <1528e1b4-3dee-161b-9463-57471263b5a8@linux.intel.com>
 <6a57b701-99a2-3917-3879-bc8141dca9d4@linux.intel.com>
Organization: Intel OTC
Message-ID: <c5e4723a-3136-e017-5997-77f56c4c27c1@linux.intel.com>
Date:   Thu, 16 Apr 2020 22:41:57 +0800
User-Agent: Mozilla/5.0 (Windows NT 10.0; WOW64; rv:68.0) Gecko/20100101
 Thunderbird/68.7.0
MIME-Version: 1.0
In-Reply-To: <6a57b701-99a2-3917-3879-bc8141dca9d4@linux.intel.com>
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Transfer-Encoding: 8bit
Content-Language: en-US
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

Ping.

On 2020/4/8 22:04, Like Xu wrote:
> Hi Paolo,
>
> Could you please take a look at this patch?
> If there is anything needs to be improved, please let me know.
>
> Thanks,
> Like Xu
>
> On 2020/3/26 20:47, Like Xu wrote:
>> Anyone to help review this change?
>>
>> Thanks,
>> Like Xu
>>
>> On 2020/3/17 16:14, Like Xu wrote:
>>> The cost of perf_event_period() is unstable, and when the guest samples
>>> multiple events, the overhead increases dramatically (5378 ns on E5-2699).
>>>
>>> For a non-running counter, the effective time of the new period is when
>>> its corresponding enable bit is enabled. Calling perf_event_period()
>>> in advance is superfluous. For a running counter, it's safe to delay the
>>> effective time until the KVM_REQ_PMU event is handled. If there are
>>> multiple perf_event_period() calls before handling KVM_REQ_PMU,
>>> it helps to reduce the total cost.
>>>
>>> Signed-off-by: Like Xu <like.xu@linux.intel.com>
>>> ---
>>>   arch/x86/kvm/pmu.c           | 11 -----------
>>>   arch/x86/kvm/pmu.h           | 11 +++++++++++
>>>   arch/x86/kvm/vmx/pmu_intel.c | 10 ++++------
>>>   3 files changed, 15 insertions(+), 17 deletions(-)
>>>
>>> diff --git a/arch/x86/kvm/pmu.c b/arch/x86/kvm/pmu.c
>>> index d1f8ca57d354..527a8bb85080 100644
>>> --- a/arch/x86/kvm/pmu.c
>>> +++ b/arch/x86/kvm/pmu.c
>>> @@ -437,17 +437,6 @@ void kvm_pmu_init(struct kvm_vcpu *vcpu)
>>>       kvm_pmu_refresh(vcpu);
>>>   }
>>> -static inline bool pmc_speculative_in_use(struct kvm_pmc *pmc)
>>> -{
>>> -    struct kvm_pmu *pmu = pmc_to_pmu(pmc);
>>> -
>>> -    if (pmc_is_fixed(pmc))
>>> -        return fixed_ctrl_field(pmu->fixed_ctr_ctrl,
>>> -            pmc->idx - INTEL_PMC_IDX_FIXED) & 0x3;
>>> -
>>> -    return pmc->eventsel & ARCH_PERFMON_EVENTSEL_ENABLE;
>>> -}
>>> -
>>>   /* Release perf_events for vPMCs that have been unused for a full 
>>> time slice.  */
>>>   void kvm_pmu_cleanup(struct kvm_vcpu *vcpu)
>>>   {
>>> diff --git a/arch/x86/kvm/pmu.h b/arch/x86/kvm/pmu.h
>>> index d7da2b9e0755..cd112e825d2c 100644
>>> --- a/arch/x86/kvm/pmu.h
>>> +++ b/arch/x86/kvm/pmu.h
>>> @@ -138,6 +138,17 @@ static inline u64 get_sample_period(struct kvm_pmc 
>>> *pmc, u64 counter_value)
>>>       return sample_period;
>>>   }
>>> +static inline bool pmc_speculative_in_use(struct kvm_pmc *pmc)
>>> +{
>>> +    struct kvm_pmu *pmu = pmc_to_pmu(pmc);
>>> +
>>> +    if (pmc_is_fixed(pmc))
>>> +        return fixed_ctrl_field(pmu->fixed_ctr_ctrl,
>>> +            pmc->idx - INTEL_PMC_IDX_FIXED) & 0x3;
>>> +
>>> +    return pmc->eventsel & ARCH_PERFMON_EVENTSEL_ENABLE;
>>> +}
>>> +
>>>   void reprogram_gp_counter(struct kvm_pmc *pmc, u64 eventsel);
>>>   void reprogram_fixed_counter(struct kvm_pmc *pmc, u8 ctrl, int 
>>> fixed_idx);
>>>   void reprogram_counter(struct kvm_pmu *pmu, int pmc_idx);
>>> diff --git a/arch/x86/kvm/vmx/pmu_intel.c b/arch/x86/kvm/vmx/pmu_intel.c
>>> index 7c857737b438..20f654a0c09b 100644
>>> --- a/arch/x86/kvm/vmx/pmu_intel.c
>>> +++ b/arch/x86/kvm/vmx/pmu_intel.c
>>> @@ -263,15 +263,13 @@ static int intel_pmu_set_msr(struct kvm_vcpu 
>>> *vcpu, struct msr_data *msr_info)
>>>               if (!msr_info->host_initiated)
>>>                   data = (s64)(s32)data;
>>>               pmc->counter += data - pmc_read_counter(pmc);
>>> -            if (pmc->perf_event)
>>> -                perf_event_period(pmc->perf_event,
>>> -                          get_sample_period(pmc, data));
>>> +            if (pmc_speculative_in_use(pmc))
>>> +                kvm_make_request(KVM_REQ_PMU, vcpu);
>>>               return 0;
>>>           } else if ((pmc = get_fixed_pmc(pmu, msr))) {
>>>               pmc->counter += data - pmc_read_counter(pmc);
>>> -            if (pmc->perf_event)
>>> -                perf_event_period(pmc->perf_event,
>>> -                          get_sample_period(pmc, data));
>>> +            if (pmc_speculative_in_use(pmc))
>>> +                kvm_make_request(KVM_REQ_PMU, vcpu);
>>>               return 0;
>>>           } else if ((pmc = get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0))) {
>>>               if (data == pmc->eventsel)
>>>
>>
>