Re: [PATCH 3/5] KVM: VMX: RTIT_CTL_BRANCH_EN has no dependency on other CPUID bit

From: Xiaoyao Li <xiaoyao.li@intel.com>
To: Like Xu <like.xu.linux@gmail.com>
Cc: 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>,
	kvm@vger.kernel.org, linux-kernel@vger.kernel.org,
	Paolo Bonzini <pbonzini@redhat.com>,
	"Alexander Shishkin (hwtracing + intel_th + stm + R:perf)" 
	<alexander.shishkin@linux.intel.com>
Subject: Re: [PATCH 3/5] KVM: VMX: RTIT_CTL_BRANCH_EN has no dependency on other CPUID bit
Date: Wed, 25 Aug 2021 16:58:09 +0800	[thread overview]
Message-ID: <d3ee1f77-8e15-f24f-d55f-cea272faf55c@intel.com> (raw)
In-Reply-To: <26103eda-806a-516d-096f-c9e85552286a@gmail.com>

On 8/25/2021 4:14 PM, Like Xu wrote:
> On 25/8/2021 2:33 pm, Xiaoyao Li wrote:
>> On 8/25/2021 2:08 PM, Like Xu wrote:
>>> On 25/8/2021 12:19 pm, Xiaoyao Li wrote:
>>>> On 8/25/2021 11:30 AM, Like Xu wrote:
>>>>> +Alexander
>>>>>
>>>>> On 24/8/2021 7:07 pm, Xiaoyao Li wrote:
>>>>>> Per Intel SDM, RTIT_CTL_BRANCH_EN bit has no dependency on any CPUID
>>>>>> leaf 0x14.
>>>>>>
>>>>>> Signed-off-by: Xiaoyao Li <xiaoyao.li@intel.com>
>>>>>> ---
>>>>>>   arch/x86/kvm/vmx/vmx.c | 8 ++++----
>>>>>>   1 file changed, 4 insertions(+), 4 deletions(-)
>>>>>>
>>>>>> diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
>>>>>> index 7ed96c460661..4a70a6d2f442 100644
>>>>>> --- a/arch/x86/kvm/vmx/vmx.c
>>>>>> +++ b/arch/x86/kvm/vmx/vmx.c
>>>>>> @@ -7116,7 +7116,8 @@ static void update_intel_pt_cfg(struct 
>>>>>> kvm_vcpu *vcpu)
>>>>>>       /* Initialize and clear the no dependency bits */
>>>>>>       vmx->pt_desc.ctl_bitmask = ~(RTIT_CTL_TRACEEN | RTIT_CTL_OS |
>>>>>> -            RTIT_CTL_USR | RTIT_CTL_TSC_EN | RTIT_CTL_DISRETC);
>>>>>> +            RTIT_CTL_USR | RTIT_CTL_TSC_EN | RTIT_CTL_DISRETC |
>>>>>> +            RTIT_CTL_BRANCH_EN);
>>>>>>       /*
>>>>>>        * If CPUID.(EAX=14H,ECX=0):EBX[0]=1 CR3Filter can be set 
>>>>>> otherwise
>>>>>> @@ -7134,12 +7135,11 @@ static void update_intel_pt_cfg(struct 
>>>>>> kvm_vcpu *vcpu)
>>>>>>                   RTIT_CTL_CYC_THRESH | RTIT_CTL_PSB_FREQ);
>>>>>>       /*
>>>>>> -     * If CPUID.(EAX=14H,ECX=0):EBX[3]=1 MTCEn BranchEn and
>>>>>> -     * MTCFreq can be set
>>>>>> +     * If CPUID.(EAX=14H,ECX=0):EBX[3]=1 MTCEn and MTCFreq can be 
>>>>>> set
>>>>>
>>>>> If CPUID.(EAX=14H,ECX=0):EBX[3]=1,
>>>>>
>>>>>      "indicates support of MTC timing packet and suppression of 
>>>>> COFI-based packets."
>>>>
>>>> I think it's a mistake of SDM in CPUID instruction.
>>>>
>>>> If you read 31.3.1, table 31-11 of SDM 325462-075US,
>>>>
>>>> It just says CPUID(0x14, 0):EBX[3]: MTC supprted.
>>>> It doesn't talk anything about COFI packets suppression.
>>>>
>>>> Further as below.
>>>>
>>>>> Per 31.2.5.4 Branch Enable (BranchEn),
>>>>>
>>>>>      "If BranchEn is not set, then relevant COFI packets (TNT, 
>>>>> TIP*, FUP, MODE.*) are suppressed."
>>>>>
>>>>> I think if the COFI capability is suppressed, the software can't 
>>>>> set the BranchEn bit, right ?
>>>>
>>>> Based on your understanding, isn't it that
>>>>
>>>> 1. if CPUID.(EAX=14H,ECX=0):EBX[3]=0, it doesn't support 
>>>> "suppression of COFI-based packets".
>>>> 2. if it doesn't support "suppression of COFI-based packets", then 
>>>> it doens't support "If BranchEn is not set, then relevant COFI 
>>>> packets (TNT, TIP*, FUP, MODE.*) are suppressed", i.e. BranchEn must 
>>>> be 1.
>>>
>>> That's it.
>>>
>>>>
>>>> Anyway, I think it's just a mistake on CPUID instruction document of 
>>>> SDM.
>>>
>>> Is this an ambiguity rather than a mistake ?
>>>
>>>>
>>>> CPUD.(EAX=14H,ECX=0):EBX[3] should only indicates the MTC support.
>>>
>>> Please do not make assertions that you do not confirm with hw.
>>>
>>>>
>>>> BranchEn should be always supported if PT is available. Per "31.2.7.2 
>>>
>>> Check d35869ba348d3f1ff3e6d8214fe0f674bb0e404e.
>>
>> This commit shows BranchEn is supported on BDW, and must be enabled on 
>> BDW. This doesn't conflict the description above that BranchEn should 
>> be always supported.
> 
> Per Vol. 4 Table 2-34. Additional MSRs Common to Processors Based the
> Broadwell Microarchitectures, the BranchEn bit 13 is:
> 
>      "Reserved; writing 0 will #GP if also setting TraceEn"
> 
> on the Intel® Core™ M Processors.
> 
> My point is that we, especially software developers from hardware vendors,
> should really focus on real hardware and fix real problems.

Isn't this patch fixing real problem? Without it, it forbids guest to 
enable BranchEn if PT_MTC_cap not exposed to guest.