Re: [PATCH v11 2/9] KVM: VMX: Set guest CET MSRs per KVM and host configuration

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

On Thu, Apr 23, 2020 at 09:27:49AM -0700, Sean Christopherson wrote:
> On Thu, Mar 26, 2020 at 04:18:39PM +0800, Yang Weijiang wrote:
> > @@ -3033,6 +3033,13 @@ void vmx_set_cr3(struct kvm_vcpu *vcpu, unsigned long cr3)
> >  		vmcs_writel(GUEST_CR3, guest_cr3);
> >  }
> >  
> > +static bool is_cet_mode_allowed(struct kvm_vcpu *vcpu, u32 mode_mask)
> 
> CET itself isn't a mode.  And since this ends up being an inner helper for
> is_cet_supported(), I think __is_cet_supported() would be the way to go.
> 
> Even @mode_mask is a bit confusing without the context of it being kernel
> vs. user.  The callers are very readable, e.g. I'd much prefer passing the
> mask as opposed to doing 'bool kernel'.  Maybe s/mode_mask/cet_mask?  That
> doesn't exactly make things super clear, but at least the reader knows the
> mask is for CET features.
Make sense, will change it.

> 
> > +{
> > +	return ((supported_xss & mode_mask) &&
> > +		(guest_cpuid_has(vcpu, X86_FEATURE_SHSTK) ||
> > +		guest_cpuid_has(vcpu, X86_FEATURE_IBT)));
> > +}
> > +
> >  int vmx_set_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
> >  {
> >  	struct vcpu_vmx *vmx = to_vmx(vcpu);
> > @@ -7064,6 +7071,35 @@ static void update_intel_pt_cfg(struct kvm_vcpu *vcpu)
> >  		vmx->pt_desc.ctl_bitmask &= ~(0xfULL << (32 + i * 4));
> >  }
> >  
> > +static void vmx_update_intercept_for_cet_msr(struct kvm_vcpu *vcpu)
> > +{
> > +	struct vcpu_vmx *vmx = to_vmx(vcpu);
> > +	unsigned long *msr_bitmap = vmx->vmcs01.msr_bitmap;
> > +	bool flag;
> 
> Maybe s/flag/incpt or something to make it more obvious that the bool is
> true if we want to intercept?  vmx_set_intercept_for_msr()s's @value isn't
> any better :-/.
I prefer using incpt now ;-) 
> > +
> > +	flag = !is_cet_mode_allowed(vcpu, XFEATURE_MASK_CET_USER);
> > +	/*
> > +	 * U_CET is required for USER CET, and U_CET, PL3_SPP are bound as
> > +	 * one component and controlled by IA32_XSS[bit 11].
> > +	 */
> > +	vmx_set_intercept_for_msr(msr_bitmap, MSR_IA32_U_CET, MSR_TYPE_RW, flag);
> > +	vmx_set_intercept_for_msr(msr_bitmap, MSR_IA32_PL3_SSP, MSR_TYPE_RW, flag);
> > +
> > +	flag = !is_cet_mode_allowed(vcpu, XFEATURE_MASK_CET_KERNEL);
> > +	/*
> > +	 * S_CET is required for KERNEL CET, and PL0_SSP ... PL2_SSP are
> > +	 * bound as one component and controlled by IA32_XSS[bit 12].
> > +	 */
> > +	vmx_set_intercept_for_msr(msr_bitmap, MSR_IA32_S_CET, MSR_TYPE_RW, flag);
> > +	vmx_set_intercept_for_msr(msr_bitmap, MSR_IA32_PL0_SSP, MSR_TYPE_RW, flag);
> > +	vmx_set_intercept_for_msr(msr_bitmap, MSR_IA32_PL1_SSP, MSR_TYPE_RW, flag);
> > +	vmx_set_intercept_for_msr(msr_bitmap, MSR_IA32_PL2_SSP, MSR_TYPE_RW, flag);
> > +
> > +	flag |= !guest_cpuid_has(vcpu, X86_FEATURE_SHSTK);
> > +	/* SSP_TAB is only available for KERNEL SHSTK.*/
> > +	vmx_set_intercept_for_msr(msr_bitmap, MSR_IA32_INT_SSP_TAB, MSR_TYPE_RW, flag);
> > +}
> > +
> >  static void vmx_cpuid_update(struct kvm_vcpu *vcpu)
> >  {
> >  	struct vcpu_vmx *vmx = to_vmx(vcpu);
> > @@ -7102,6 +7138,10 @@ static void vmx_cpuid_update(struct kvm_vcpu *vcpu)
> >  			vmx_set_guest_msr(vmx, msr, enabled ? 0 : TSX_CTRL_RTM_DISABLE);
> >  		}
> >  	}
> > +
> > +	if (guest_cpuid_has(vcpu, X86_FEATURE_SHSTK) ||
> > +	    guest_cpuid_has(vcpu, X86_FEATURE_IBT))
> > +		vmx_update_intercept_for_cet_msr(vcpu);
> 
> This is wrong, it will miss the case where userspace double configures CPUID
> and goes from CET=1 to CET=0.  This should instead be:
> 
> 	if (supported_xss & (XFEATURE_MASK_CET_KERNEL | XFEATURE_MASK_CET_USER))
> 		vmx_update_intercept_for_cet_msr(vcpu);
> 
> >  }
Here CET=1/0, did you mean the CET bit in XSS or CR4.CET? If it's the
former, then it's OK for me.

> >  
> >  static __init void vmx_set_cpu_caps(void)
> > -- 
> > 2.17.2
> >