Re: [PATCH v12 07/11] KVM: vmx/pmu: Unmask LBR fields in the MSR_IA32_DEBUGCTLMSR emualtion

["Xu, Like" <like.xu@intel.com>]

Hi Sean,

First of all, are you going to queue the LBR patch series in your tree
considering the host perf patches have already queued in Peter's tree ?

On 2020/7/8 4:21, Sean Christopherson wrote:
> On Sat, Jun 13, 2020 at 05:42:50PM +0800, Xu, Like wrote:
>> On 2020/6/13 17:14, Xiaoyao Li wrote:
>>> On 6/13/2020 4:09 PM, Like Xu wrote:
>>>> When the LBR feature is reported by the vmx_get_perf_capabilities(),
>>>> the LBR fields in the [vmx|vcpu]_supported debugctl should be unmasked.
>>>>
>>>> The debugctl msr is handled separately in vmx/svm and they're not
>>>> completely identical, hence remove the common msr handling code.
> I would prefer to put the "remove DEBUGCTRL handling from common x86" in a
> separate patch.  Without digging into SVM, it's not obvious that dropping
> MSR_IA32_DEBUGCTLMSR from kvm_set_msr_common() is a nop for SVM.
Sure, I'll do it in a separate patch.
>
>>>> Signed-off-by: Like Xu <like.xu@linux.intel.com>
>>>> ---
>>>>    arch/x86/kvm/vmx/capabilities.h | 12 ++++++++++++
>>>>    arch/x86/kvm/vmx/pmu_intel.c    | 19 +++++++++++++++++++
>>>>    arch/x86/kvm/x86.c              | 13 -------------
>>>>    3 files changed, 31 insertions(+), 13 deletions(-)
>>>>
>>>> diff --git a/arch/x86/kvm/vmx/capabilities.h
>>>> b/arch/x86/kvm/vmx/capabilities.h
>>>> index b633a90320ee..f6fcfabb1026 100644
>>>> --- a/arch/x86/kvm/vmx/capabilities.h
>>>> +++ b/arch/x86/kvm/vmx/capabilities.h
>>>> @@ -21,6 +21,8 @@ extern int __read_mostly pt_mode;
>>>>    #define PMU_CAP_FW_WRITES    (1ULL << 13)
>>>>    #define PMU_CAP_LBR_FMT        0x3f
>>>>    +#define DEBUGCTLMSR_LBR_MASK        (DEBUGCTLMSR_LBR |
>>>> DEBUGCTLMSR_FREEZE_LBRS_ON_PMI)
>>>> +
>>>>    struct nested_vmx_msrs {
>>>>        /*
>>>>         * We only store the "true" versions of the VMX capability MSRs. We
>>>> @@ -387,4 +389,14 @@ static inline u64 vmx_get_perf_capabilities(void)
>>>>        return perf_cap;
>>>>    }
>>>>    +static inline u64 vmx_get_supported_debugctl(void)
>>>> +{
>>>> +    u64 val = 0;
>>>> +
>>>> +    if (vmx_get_perf_capabilities() & PMU_CAP_LBR_FMT)
>>>> +        val |= DEBUGCTLMSR_LBR_MASK;
>>>> +
>>>> +    return val;
>>>> +}
>>>> +
>>>>    #endif /* __KVM_X86_VMX_CAPS_H */
>>>> diff --git a/arch/x86/kvm/vmx/pmu_intel.c b/arch/x86/kvm/vmx/pmu_intel.c
>>>> index a953c7d633f6..d92e95b64c74 100644
>>>> --- a/arch/x86/kvm/vmx/pmu_intel.c
>>>> +++ b/arch/x86/kvm/vmx/pmu_intel.c
>>>> @@ -187,6 +187,7 @@ static bool intel_is_valid_msr(struct kvm_vcpu
>>>> *vcpu, u32 msr)
>>>>        case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
>>>>            ret = pmu->version > 1;
>>>>            break;
>>>> +    case MSR_IA32_DEBUGCTLMSR:
>>>>        case MSR_IA32_PERF_CAPABILITIES:
>>>>            ret = 1;
>>>>            break;
>>>> @@ -237,6 +238,9 @@ static int intel_pmu_get_msr(struct kvm_vcpu *vcpu,
>>>> struct msr_data *msr_info)
>>>>                return 1;
>>>>            msr_info->data = vcpu->arch.perf_capabilities;
>>>>            return 0;
>>>> +    case MSR_IA32_DEBUGCTLMSR:
>>>> +        msr_info->data = vmcs_read64(GUEST_IA32_DEBUGCTL);
>>> Can we put the emulation of MSR_IA32_DEBUGCTLMSR in vmx_{get/set})_msr().
>>> AFAIK, MSR_IA32_DEBUGCTLMSR is not a pure PMU related MSR that there is
>>> bit 2 to enable #DB for bus lock.
>> We already have "case MSR_IA32_DEBUGCTLMSR" handler in the vmx_set_msr()
>> and you may apply you bus lock changes in that handler.
> Hrm, but that'd be weird dependency as vmx_set_msr() would need to check for
> #DB bus lock support but not actually write GUEST_IA32_DEBUGCTL, or we'd end
> up writing it twice when both bus lock and LBR are supported.
Yes, you're right about the multiple writes on GUEST_IA32_DEBUGCTL.

I'll move the handler to vmx_set/get_msr() for other DEBUGCTL users.
>
> I don't see anything in the series that takes action on writes to
> MSR_IA32_DEBUGCTLMSR beyond updating the VMCS, i.e. AFAICT there isn't any
> reason to call into the PMU, VMX can simply query vmx_get_perf_capabilities()
> to check if it's legal to enable DEBUGCTLMSR_LBR_MASK.
There's a gap to enable DEBUGCTLMSR_LBR_MASK.

The vmx_get_perf_capabilities() is queried per-KVM while
the vmx_get_supported_debugctl() is queried per-guest.

>
> A question for both LBR and bus lock: would it make sense to cache the
> guest's value in vcpu_vmx so that querying the guest value doesn't require
> a VMREAD?  I don't have a good feel for how frequently it would be accessed.
I'm OK with the cached value for this field and AFAIK,
it will benefit the legacy_freezing_lbrs_on_pmi emulation
if the VMREAD is heavier than normal cache/mem touch.

>
>>>> +        return 0;
>>>>        default:
>>>>            if ((pmc = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0)) ||
>>>>                (pmc = get_gp_pmc(pmu, msr, MSR_IA32_PMC0))) {
>>>> @@ -282,6 +286,16 @@ static inline bool lbr_is_compatible(struct
>>>> kvm_vcpu *vcpu)
>>>>        return true;
>>>>    }
>>>>    +static inline u64 vcpu_get_supported_debugctl(struct kvm_vcpu *vcpu)
>>>> +{
>>>> +    u64 debugctlmsr = vmx_get_supported_debugctl();
>>>> +
>>>> +    if (!lbr_is_enabled(vcpu))
>>>> +        debugctlmsr &= ~DEBUGCTLMSR_LBR_MASK;
>>>> +
>>>> +    return debugctlmsr;
>>>> +}
>>>> +
>>>>    static int intel_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data
>>>> *msr_info)
>>>>    {
>>>>        struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
>>>> @@ -336,6 +350,11 @@ static int intel_pmu_set_msr(struct kvm_vcpu *vcpu,
>>>> struct msr_data *msr_info)
>>>>            }
>>>>            vcpu->arch.perf_capabilities = data;
>>>>            return 0;
>>>> +    case MSR_IA32_DEBUGCTLMSR:
>>>> +        if (data & ~vcpu_get_supported_debugctl(vcpu))
>>>> +            return 1;
>>>> +        vmcs_write64(GUEST_IA32_DEBUGCTL, data);
>>>> +        return 0;
>>>>        default:
>>>>            if ((pmc = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0)) ||
>>>>                (pmc = get_gp_pmc(pmu, msr, MSR_IA32_PMC0))) {
>>>> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
>>>> index 00c88c2f34e4..56f275eb4554 100644
>>>> --- a/arch/x86/kvm/x86.c
>>>> +++ b/arch/x86/kvm/x86.c
>>>> @@ -2840,18 +2840,6 @@ int kvm_set_msr_common(struct kvm_vcpu *vcpu,
>>>> struct msr_data *msr_info)
>>>>                return 1;
>>>>            }
>>>>            break;
>>>> -    case MSR_IA32_DEBUGCTLMSR:
>>>> -        if (!data) {
>>>> -            /* We support the non-activated case already */
>>>> -            break;
>>>> -        } else if (data & ~(DEBUGCTLMSR_LBR | DEBUGCTLMSR_BTF)) {
>>> So after this patch, guest trying to set bit DEBUGCTLMSR_BTF will get a
>>> #GP instead of being ignored and printing a log in kernel.
>>>
>> Since the BTF is not implemented on the KVM at all,
>> I do propose not left this kind of dummy thing in the future KVM code.
>>
>> Let's see if Netware or any BTF user will complain about this change.
> If you want to drop that behavior it needs be done in a separate patch.
> Personally I don't see the point in doing so, it's a trivial amount of code
> in KVM and there's no harm in dropping the bits on write.
No harm in dropping the bits on write ? Interesting.
I may keep the semantics unchanged for LBR patches and make it as separate 
proposal.

Thanks,
Like Xu