Re: [PATCH v6 07/16] KVM: x86/pmu: Reprogram PEBS event to emulate guest PEBS counter

From: "Xu, Like" <like.xu@intel.com>
To: Peter Zijlstra <peterz@infradead.org>, Andi Kleen <ak@linux.intel.com>
Cc: Like Xu <like.xu@linux.intel.com>,
	Paolo Bonzini <pbonzini@redhat.com>,
	Borislav Petkov <bp@alien8.de>,
	Sean Christopherson <seanjc@google.com>,
	Vitaly Kuznetsov <vkuznets@redhat.com>,
	Wanpeng Li <wanpengli@tencent.com>,
	Jim Mattson <jmattson@google.com>, Joerg Roedel <joro@8bytes.org>,
	weijiang.yang@intel.com, Kan Liang <kan.liang@linux.intel.com>,
	wei.w.wang@intel.com, eranian@google.com,
	liuxiangdong5@huawei.com, linux-kernel@vger.kernel.org,
	x86@kernel.org, kvm@vger.kernel.org
Subject: Re: [PATCH v6 07/16] KVM: x86/pmu: Reprogram PEBS event to emulate guest PEBS counter
Date: Tue, 18 May 2021 21:15:58 +0800	[thread overview]
Message-ID: <852ab586-2438-c7fc-c41d-0862e2f1b7ca@intel.com> (raw)
In-Reply-To: <YKN/DVNt847iEctd@hirez.programming.kicks-ass.net>

On 2021/5/18 16:47, Peter Zijlstra wrote:
> On Mon, May 17, 2021 at 07:44:15AM -0700, Andi Kleen wrote:
>> On 5/17/2021 1:39 AM, Peter Zijlstra wrote:
>>> On Tue, May 11, 2021 at 10:42:05AM +0800, Like Xu wrote:
>>>> +	if (pebs) {
>>>> +		/*
>>>> +		 * The non-zero precision level of guest event makes the ordinary
>>>> +		 * guest event becomes a guest PEBS event and triggers the host
>>>> +		 * PEBS PMI handler to determine whether the PEBS overflow PMI
>>>> +		 * comes from the host counters or the guest.
>>>> +		 *
>>>> +		 * For most PEBS hardware events, the difference in the software
>>>> +		 * precision levels of guest and host PEBS events will not affect
>>>> +		 * the accuracy of the PEBS profiling result, because the "event IP"
>>>> +		 * in the PEBS record is calibrated on the guest side.
>>>> +		 */
>>>> +		attr.precise_ip = 1;
>>>> +	}
>>> You've just destroyed precdist, no?
>> precdist can mean multiple things:
>>
>> - Convert cycles to the precise INST_RETIRED event. That is not meaningful
>> for virtualization because "cycles" doesn't exist, just the raw events.
>>
>> - For GLC+ and TNT+ it will force the event to a specific counter that is
>> more precise. This would be indeed "destroyed", but right now the patch kit
>> only supports Icelake which doesn't support that anyways.
>>
>> So I think the code is correct for now, but will need to be changed for
>> later CPUs. Should perhaps fix the comment though to discuss this.
> OK, can we then do a better comment that explains *why* this is correct
> now and what needs help later?
>
> Because IIUC the only reason it is correct now is because:
>
>   - we only support ICL
>
>     * and ICL has pebs_format>=2, so {1,2} are the same
>     * and ICL doesn't have precise_ip==3 support
>
>   - Other hardware (GLC+, TNT+) that could possibly care here
>     is unsupported atm. but needs changes.
>
> None of which is actually mentioned in that comment it does have.

Hi Andi & Peter,

By "precdist", do you mean the"Precise Distribution of Instructions Retired 
(PDIR) Facility"?

The SDM says Ice Lake Microarchitecture does support PEBS-PDIR on 
IA32_FIXED0 only.
And this patch kit enables it in the patch 0011, please take a look.

Or do I miss something about precdist on ICL ?

Thanks,
Like Xu