Re: [PATCH RFC 2/3] x86/kvm/hyper-v: move VMX controls sanitization out of nested_enable_evmcs()

[Vitaly Kuznetsov <vkuznets@redhat.com>]

Sean Christopherson <sean.j.christopherson@intel.com> writes:

> On Wed, Jan 22, 2020 at 04:08:55PM +0100, Vitaly Kuznetsov wrote:
>> Paolo Bonzini <pbonzini@redhat.com> writes:
>> 
>> > On 22/01/20 06:47, Sean Christopherson wrote:
>> >>> Yes, it most likely is and it would be nice if Microsoft fixed it, but I
>> >>> guess we're stuck with it for existing Windows versions.  Well, for one
>> >>> we found a bug in Hyper-V and not the converse. :)
>> >>>
>> >>> There is a problem with this approach, in that we're stuck with it
>> >>> forever due to live migration.  But I guess if in the future eVMCS v2
>> >>> adds an apic_address field we can limit the hack to eVMCS v1.  Another
>> >>> possibility is to use the quirks mechanism but it's overkill for now.
>> >>>
>> >>> Unless there are objections, I plan to apply these patches.
>> >> Doesn't applying this patch contradict your earlier opinion?  This patch
>> >> would still hide the affected controls from the guest because the host
>> >> controls enlightened_vmcs_enabled.
>> >
>> > It does.  Unfortunately the key sentence is "we're stuck with it for
>> > existing Windows versions". :(
>
> Ah, I didn't understand what "it" referred to :-)
>
>> >> Rather than update vmx->nested.msrs or filter vmx_get_msr(), what about
>> >> manually adding eVMCS consistency checks on the disallowed bits and handle
>> >> SECONDARY_EXEC_VIRTUALIZE_APIC_ACCESSES as a one-off case by simply
>> >> clearing it from the eVMCS?  Or alternatively, squashing all the disallowed
>> >> bits.
>> >
>> > Hmm, that is also a possibility.  It's a very hacky one, but I guess
>> > adding APIC virtualization to eVMCS would require bumping the version to
>> > 2.  Vitaly, what do you think?
>> 
>> As I already replied to Sean I like the idea to filter out unsupported
>> controls from eVMCS but unfortunately it doesn't work: Hyper-V actually
>> expects APIC virtualization to work when it enables
>> SECONDARY_EXEC_VIRTUALIZE_APIC_ACCESSES (I have no idea how without
>> apic_access_addr field but). I checked and at least Hyper-V 2016 doesn't
>> boot (when >1 vCPU).
>
> Nice.
>
> I still don't see what we gain from applying this patch.  Once eVMCS is
> enabled by userspace, which presumably happens before the guest is launched,
> the guest will see the eVMCS-unfriendly controls as being unsupported, both
> for eVMCS and regular VMCS.  AFAICT, we're adding a fairly ugly hack to KVM
> just so that KVM can lie to userspace about what controls will be exposed to
> the guest.
>
> Can we extend the API to use cap->args[1] to control whether or not the
> unsupported controls are removed from vmx->nested.msrs?  Userspace could
> pass '1' to leave the controls untouched and then surgically hide the
> controls that the guest is too dumb to know it shouldn't use by writing the
> appropriate MSRs.  Assuming existing userspace is expected/required to zero
> out args[1..3], this would be fully backwards compatible.

Yes, in case we're back to the idea to filter things out in QEMU we can
do this. What I don't like is that every other userspace which decides
to enable eVMCS will have to perform the exact same surgery as in case
it sets allow_unsupported_controls=0 it'll have to know (hardcode) the
filtering (or KVM_SET_MSRS will fail) and in case it opts for
allow_unsupported_controls=1 Windows guests just won't boot without the
filtering.

It seems to be 1:1, eVMCSv1 requires the filter.

>
>
> diff --git a/arch/x86/kvm/vmx/evmcs.c b/arch/x86/kvm/vmx/evmcs.c
> index 72359709cdc1..241a769be738 100644
> --- a/arch/x86/kvm/vmx/evmcs.c
> +++ b/arch/x86/kvm/vmx/evmcs.c
> @@ -346,8 +346,8 @@ uint16_t nested_get_evmcs_version(struct kvm_vcpu *vcpu)
>         return 0;
>  }
>
> -int nested_enable_evmcs(struct kvm_vcpu *vcpu,
> -                       uint16_t *vmcs_version)
> +int nested_enable_evmcs(struct kvm_vcpu *vcpu, uint16_t *vmcs_version,
> +                       bool allow_unsupported_controls)

Personally, I'd call it 'keep_unsupported_controls'.

>  {
>         struct vcpu_vmx *vmx = to_vmx(vcpu);
>         bool evmcs_already_enabled = vmx->nested.enlightened_vmcs_enabled;
> @@ -358,7 +358,7 @@ int nested_enable_evmcs(struct kvm_vcpu *vcpu,
>                 *vmcs_version = nested_get_evmcs_version(vcpu);
>
>         /* We don't support disabling the feature for simplicity. */
> -       if (evmcs_already_enabled)
> +       if (evmcs_already_enabled || allow_unsupported_controls)
>                 return 0;
>
>         vmx->nested.msrs.pinbased_ctls_high &= ~EVMCS1_UNSUPPORTED_PINCTRL;
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index 0cccc52e2d0a..5e1b8d51277b 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -4005,7 +4005,8 @@ static int kvm_vcpu_ioctl_enable_cap(struct kvm_vcpu *vcpu,
>         case KVM_CAP_HYPERV_ENLIGHTENED_VMCS:
>                 if (!kvm_x86_ops->nested_enable_evmcs)
>                         return -ENOTTY;
> -               r = kvm_x86_ops->nested_enable_evmcs(vcpu, &vmcs_version);
> +               r = kvm_x86_ops->nested_enable_evmcs(vcpu, &vmcs_version,
> +                                                    cap->args[1]);
>                 if (!r) {
>                         user_ptr = (void __user *)(uintptr_t)cap->args[0];
>                         if (copy_to_user(user_ptr, &vmcs_version,
>

-- 
Vitaly