Re: [PATCH v5 07/34] KVM: SVM: Add required changes to support intercepts under SEV-ES

[Tom Lendacky <thomas.lendacky@amd.com>]

On 12/14/20 9:33 AM, Paolo Bonzini wrote:
> On 10/12/20 18:09, Tom Lendacky wrote:
>> @@ -2797,7 +2838,27 @@ static int svm_set_msr(struct kvm_vcpu *vcpu,
>> struct msr_data *msr)
>>     static int wrmsr_interception(struct vcpu_svm *svm)
>>   {
>> -    return kvm_emulate_wrmsr(&svm->vcpu);
>> +    u32 ecx;
>> +    u64 data;
>> +
>> +    if (!sev_es_guest(svm->vcpu.kvm))
>> +        return kvm_emulate_wrmsr(&svm->vcpu);
>> +
>> +    ecx = kvm_rcx_read(&svm->vcpu);
>> +    data = kvm_read_edx_eax(&svm->vcpu);
>> +    if (kvm_set_msr(&svm->vcpu, ecx, data)) {
>> +        trace_kvm_msr_write_ex(ecx, data);
>> +        ghcb_set_sw_exit_info_1(svm->ghcb, 1);
>> +        ghcb_set_sw_exit_info_2(svm->ghcb,
>> +                    X86_TRAP_GP |
>> +                    SVM_EVTINJ_TYPE_EXEPT |
>> +                    SVM_EVTINJ_VALID);
>> +        return 1;
>> +    }
>> +
>> +    trace_kvm_msr_write(ecx, data);
>> +
>> +    return kvm_skip_emulated_instruction(&svm->vcpu);
>>   }
>>     static int msr_interception(struct vcpu_svm *svm)
> 
> This code duplication is ugly, and does not work with userspace MSR
> filters too.

Agree and I missed that the userspace MSR support went in.

> 
> But we can instead trap the completion of the MSR read/write to use
> ghcb_set_sw_exit_info_1 instead of kvm_inject_gp, with a callback like
> 
> static int svm_complete_emulated_msr(struct kvm_vcpu *vcpu, int err)
> {
>         if (!sev_es_guest(svm->vcpu.kvm) || !err)
>                 return kvm_complete_insn_gp(&svm->vcpu, err);
> 
>         ghcb_set_sw_exit_info_1(svm->ghcb, 1);
>         ghcb_set_sw_exit_info_2(svm->ghcb,
>                                 X86_TRAP_GP |
>                                 SVM_EVTINJ_TYPE_EXEPT |
>                                 SVM_EVTINJ_VALID);
>         return 1;
> }

If we use the kvm_complete_insn_gp() we lose the tracing and it needs to
be able to deal with read completion setting the registers. It also needs
to deal with both kvm_emulate_rdmsr/wrmsr() when not bouncing to
userspace. Let me take a shot at covering all the cases and see what I can
come up with.

I noticed that the userspace completion path doesn't have tracing
invocations, trace_kvm_msr_read/write_ex() or trace_kvm_msr_read/write(),
is that by design?

> 
> 
> ...
>     .complete_emulated_msr = svm_complete_emulated_msr,
> 
>> @@ -2827,7 +2888,14 @@ static int interrupt_window_interception(struct
>> vcpu_svm *svm)
>>   static int pause_interception(struct vcpu_svm *svm)
>>   {
>>       struct kvm_vcpu *vcpu = &svm->vcpu;
>> -    bool in_kernel = (svm_get_cpl(vcpu) == 0);
>> +    bool in_kernel;
>> +
>> +    /*
>> +     * CPL is not made available for an SEV-ES guest, so just set
>> in_kernel
>> +     * to true.
>> +     */
>> +    in_kernel = (sev_es_guest(svm->vcpu.kvm)) ? true
>> +                          : (svm_get_cpl(vcpu) == 0);
>>         if (!kvm_pause_in_guest(vcpu->kvm))
>>           grow_ple_window(vcpu);
> 
> See below.
> 
>> @@ -3273,6 +3351,13 @@ bool svm_interrupt_blocked(struct kvm_vcpu *vcpu)
>>       struct vcpu_svm *svm = to_svm(vcpu);
>>       struct vmcb *vmcb = svm->vmcb;
>>   +    /*
>> +     * SEV-ES guests to not expose RFLAGS. Use the VMCB interrupt mask
>> +     * bit to determine the state of the IF flag.
>> +     */
>> +    if (sev_es_guest(svm->vcpu.kvm))
>> +        return !(vmcb->control.int_state & SVM_GUEST_INTERRUPT_MASK);
> 
> This seems wrong, you have to take into account SVM_INTERRUPT_SHADOW_MASK
> as well.  Also, even though GIF is not really used by SEV-ES guests, I
> think it's nicer to put this check afterwards.
> 
> That is:
> 
> diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
> index 4372e45c8f06..2dd9c9698480 100644
> --- a/arch/x86/kvm/svm/svm.c
> +++ b/arch/x86/kvm/svm/svm.c
> @@ -3247,7 +3247,14 @@ bool svm_interrupt_blocked(struct kvm_vcpu *vcpu)
>      if (!gif_set(svm))
>          return true;
> 
> -    if (is_guest_mode(vcpu)) {
> +    if (sev_es_guest(svm->vcpu.kvm)) {
> +        /*
> +         * SEV-ES guests to not expose RFLAGS. Use the VMCB interrupt mask
> +         * bit to determine the state of the IF flag.
> +         */
> +        if (!(vmcb->control.int_state & SVM_GUEST_INTERRUPT_MASK))
> +            return true;
> +    } else if (is_guest_mode(vcpu)) {
>          /* As long as interrupts are being delivered...  */
>          if ((svm->nested.ctl.int_ctl & V_INTR_MASKING_MASK)
>              ? !(svm->nested.hsave->save.rflags & X86_EFLAGS_IF)
> 
> 

Yup, I'll make that change.

> 
>>       if (!gif_set(svm))
>>           return true;
>>   @@ -3458,6 +3543,12 @@ static void svm_complete_interrupts(struct
>> vcpu_svm *svm)
>>           svm->vcpu.arch.nmi_injected = true;
>>           break;
>>       case SVM_EXITINTINFO_TYPE_EXEPT:
>> +        /*
>> +         * Never re-inject a #VC exception.
>> +         */
>> +        if (vector == X86_TRAP_VC)
>> +            break;
>> +
>>           /*
>>            * In case of software exceptions, do not reinject the vector,
>>            * but re-execute the instruction instead. Rewind RIP first
>> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
>> index a3fdc16cfd6f..b6809a2851d2 100644
>> --- a/arch/x86/kvm/x86.c
>> +++ b/arch/x86/kvm/x86.c
>> @@ -4018,7 +4018,7 @@ void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
>>   {
>>       int idx;
>>   -    if (vcpu->preempted)
>> +    if (vcpu->preempted && !vcpu->arch.guest_state_protected)
>>           vcpu->arch.preempted_in_kernel = !kvm_x86_ops.get_cpl(vcpu);
> 
> This has to be true, otherwise no directed yield will be done at all:
> 
>     if (READ_ONCE(vcpu->preempted) && yield_to_kernel_mode &&
>         !kvm_arch_vcpu_in_kernel(vcpu))
>         continue;
> 
> Or more easily, just use in_kernel == false in pause_interception, like
> 
> +    /*
> +     * CPL is not made available for an SEV-ES guest, therefore
> +     * vcpu->arch.preempted_in_kernel can never be true.  Just
> +     * set in_kernel to false as well.
> +     */
> +    in_kernel = !sev_es_guest(svm->vcpu.kvm) && svm_get_cpl(vcpu) == 0;

Sounds good, I'll make that change.

> 
>>         /*
>> @@ -8161,7 +8161,9 @@ static void post_kvm_run_save(struct kvm_vcpu *vcpu)
>>   {
>>       struct kvm_run *kvm_run = vcpu->run;
>>   -    kvm_run->if_flag = (kvm_get_rflags(vcpu) & X86_EFLAGS_IF) != 0;
>> +    kvm_run->if_flag = (vcpu->arch.guest_state_protected)
>> +        ? kvm_arch_interrupt_allowed(vcpu)
>> +        : (kvm_get_rflags(vcpu) & X86_EFLAGS_IF) != 0;
> 
> Here indeed you only want the interrupt allowed bit, not the interrupt
> window.  But we can just be bold and always set it to true.
> 
> - for userspace irqchip, kvm_run->ready_for_interrupt_injection is set
> just below and it will always be false if kvm_arch_interrupt_allowed is false
> 
> - for in-kernel APIC, if_flag is documented to be invalid (though it
> actually is valid).  For split irqchip, they can just use
> kvm_run->ready_for_interrupt_injection; for entirely in-kernel interrupt
> handling, userspace does not need if_flag at all.

Ok, I'll make that change.

Thanks,
Tom

> 
> Paolo
> 
>>       kvm_run->flags = is_smm(vcpu) ? KVM_RUN_X86_SMM : 0;
>>       kvm_run->cr8 = kvm_get_cr8(vcpu);
>>       kvm_run->apic_base = kvm_get_apic_base(vcpu);
>>
> 
>