Re: [PATCH v8 4/7] KVM: VMX: Load CET states on vmentry/vmexit

[Yang Weijiang <weijiang.yang@intel.com>]

On Tue, Dec 10, 2019 at 01:23:05PM -0800, Sean Christopherson wrote:
> On Fri, Nov 01, 2019 at 04:52:19PM +0800, Yang Weijiang wrote:
> > @@ -2834,6 +2837,9 @@ void vmx_set_cr0(struct kvm_vcpu *vcpu, unsigned long cr0)
> >  	struct vcpu_vmx *vmx = to_vmx(vcpu);
> >  	unsigned long hw_cr0;
> >  
> > +	if (!(cr0 & X86_CR0_WP) && kvm_read_cr4_bits(vcpu, X86_CR4_CET))
> > +		cr0 |= X86_CR0_WP;
> 
> Huh?  What's the interaction between CR4.CET and CR0.WP?  If there really
> is some non-standard interaction then it needs to be documented in at least
> the changelog and probably with a comment as well.
>
The processor does not allow CR4.CET to be set if CR0.WP = 0 (similarly, it does not allow CR0.WP to be
cleared while CR4.CET = 1).

> > +
> >  	hw_cr0 = (cr0 & ~KVM_VM_CR0_ALWAYS_OFF);
> >  	if (enable_unrestricted_guest)
> >  		hw_cr0 |= KVM_VM_CR0_ALWAYS_ON_UNRESTRICTED_GUEST;
> > @@ -2936,6 +2942,22 @@ static bool guest_cet_allowed(struct kvm_vcpu *vcpu, u32 feature, u32 mode)
> >  	return false;
> >  }
> >  
> > +bool is_cet_bit_allowed(struct kvm_vcpu *vcpu)
> > +{
> > +	unsigned long cr0;
> > +	bool cet_allowed;
> > +
> > +	cr0 = kvm_read_cr0(vcpu);
> > +	cet_allowed = guest_cet_allowed(vcpu, X86_FEATURE_SHSTK,
> > +					XFEATURE_MASK_CET_USER) ||
> > +		      guest_cet_allowed(vcpu, X86_FEATURE_IBT,
> > +					XFEATURE_MASK_CET_USER);
> > +	if ((cr0 & X86_CR0_WP) && cet_allowed)
> > +		return true;
> 
> So, attempting to set CR4.CET if CR0.WP=0 takes a #GP?  But attempting
> to clear CR0.WP if CR4.CET=1 is ignored?
> 
Per above words in spec., inject #GP to guest in either case?

> > +
> > +	return false;
> > +}
> > +
> >  int vmx_set_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
> >  {
> >  	struct vcpu_vmx *vmx = to_vmx(vcpu);
> > @@ -2976,6 +2998,9 @@ int vmx_set_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
> >  			return 1;
> >  	}
> >  
> > +	if ((cr4 & X86_CR4_CET) && !is_cet_bit_allowed(vcpu))
> > +		return 1;
> > +
> >  	if (vmx->nested.vmxon && !nested_cr4_valid(vcpu, cr4))
> >  		return 1;
> >  
> > @@ -3839,6 +3864,12 @@ void vmx_set_constant_host_state(struct vcpu_vmx *vmx)
> >  
> >  	if (cpu_has_load_ia32_efer())
> >  		vmcs_write64(HOST_IA32_EFER, host_efer);
> > +
> > +	if (cpu_has_load_host_cet_states_ctrl()) {
> > +		vmcs_writel(HOST_S_CET, 0);
> > +		vmcs_writel(HOST_INTR_SSP_TABLE, 0);
> > +		vmcs_writel(HOST_SSP, 0);
> > +	}
> >  }
> >  
> >  void set_cr4_guest_host_mask(struct vcpu_vmx *vmx)
> > @@ -6436,6 +6467,7 @@ static void vmx_vcpu_run(struct kvm_vcpu *vcpu)
> >  {
> >  	struct vcpu_vmx *vmx = to_vmx(vcpu);
> >  	unsigned long cr3, cr4;
> > +	bool cet_allowed;
> >  
> >  	/* Record the guest's net vcpu time for enforced NMI injections. */
> >  	if (unlikely(!enable_vnmi &&
> > @@ -6466,6 +6498,25 @@ static void vmx_vcpu_run(struct kvm_vcpu *vcpu)
> >  		vmx->loaded_vmcs->host_state.cr3 = cr3;
> >  	}
> >  
> > +	/* To be aligned with kernel code, only user mode is supported now. */
> > +	cet_allowed = guest_cet_allowed(vcpu, X86_FEATURE_SHSTK,
> > +					XFEATURE_MASK_CET_USER) ||
> > +		      guest_cet_allowed(vcpu, X86_FEATURE_IBT,
> > +					XFEATURE_MASK_CET_USER);
> > +	if (cpu_has_load_guest_cet_states_ctrl() && cet_allowed)
> > +		vmcs_set_bits(VM_ENTRY_CONTROLS,
> > +			      VM_ENTRY_LOAD_GUEST_CET_STATE);
> > +	else
> > +		vmcs_clear_bits(VM_ENTRY_CONTROLS,
> > +				VM_ENTRY_LOAD_GUEST_CET_STATE);
> > +
> > +	if (cpu_has_load_host_cet_states_ctrl() && cet_allowed)
> > +		vmcs_set_bits(VM_EXIT_CONTROLS,
> > +			      VM_EXIT_LOAD_HOST_CET_STATE);
> > +	else
> > +		vmcs_clear_bits(VM_EXIT_CONTROLS,
> > +				VM_EXIT_LOAD_HOST_CET_STATE);
> > +
> >  	cr4 = cr4_read_shadow();
> >  	if (unlikely(cr4 != vmx->loaded_vmcs->host_state.cr4)) {
> >  		vmcs_writel(HOST_CR4, cr4);
> > -- 
> > 2.17.2
> >