On Fri, Aug 28, 2020 at 01:39:39PM -0700, Jim Mattson wrote:
> On Fri, Aug 28, 2020 at 1:54 AM Chenyi Qiang <chenyi.qiang@xxxxxxxxx> wrote:
> >
> > Update the fields (i.e. VM_{ENTRY_LOAD, EXIT_CLEAR}_BNDCFGS and
> > VM_{ENTRY, EXIT}_LOAD_IA32_PERF_GLOBAL_CTRL) in
> > nested MSR_IA32_VMX_TRUE_{ENTRY, EXIT}_CTLS according to guest CPUID
> > when user space initializes the features MSRs. Regardless of the order
> > of SET_CPUID and SET_MSRS from the user space, do the update to avoid
> > MSR values overriding.
> >
> > Signed-off-by: Chenyi Qiang <chenyi.qiang@xxxxxxxxx>
> > ---
> > arch/x86/kvm/vmx/vmx.c | 6 +++++-
> > 1 file changed, 5 insertions(+), 1 deletion(-)
> >
> > diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> > index 819c185adf09..f9664ccc003b 100644
> > --- a/arch/x86/kvm/vmx/vmx.c
> > +++ b/arch/x86/kvm/vmx/vmx.c
> > @@ -345,6 +345,7 @@ static bool guest_state_valid(struct kvm_vcpu *vcpu);
> > static u32 vmx_segment_access_rights(struct kvm_segment *var);
> > static __always_inline void vmx_disable_intercept_for_msr(unsigned long *msr_bitmap,
> > u32 msr, int type);
> > +static void nested_vmx_entry_exit_ctls_update(struct kvm_vcpu *vcpu);
> >
> > void vmx_vmexit(void);
> >
> > @@ -2161,7 +2162,10 @@ static int vmx_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
> > return 1; /* they are read-only */
> > if (!nested_vmx_allowed(vcpu))
> > return 1;
> > - return vmx_set_vmx_msr(vcpu, msr_index, data);
> > + ret = vmx_set_vmx_msr(vcpu, msr_index, data);
> > + nested_vmx_pmu_entry_exit_ctls_update(vcpu);
> > + nested_vmx_entry_exit_ctls_update(vcpu);
> > + break;
>
> Now I see what you're doing. This commit should probably come before
> the previous commit, so that at no point in the series can userspace
> set VMX MSR bits that should be cleared based on the guest CPUID.
>
> There's an ABI change here: userspace may no longer get -EINVAL if it
> tries to set an illegal VMX MSR bit. Instead, some illegal bits are
> silently cleared. Moreover, these functions will potentially set VMX
> MSR bits that userspace has just asked to clear.
Can we simply remove nested_vmx_entry_exit_ctls_update() and
nested_vmx_pmu_entry_exit_ctls_update()? It's userspace's responsibility
to present a valid vCPU model to the guest, I don't see any reason to
silently tweak the VMX MSRs unless allowing the bogus config breaks KVM.
E.g. there are many more controls that are non-sensical without "native"
support for the associated feature.
