Re: [PATCH v5 3/7] KVM: X86: Expose IA32_PKRS MSR

[Sean Christopherson <seanjc@google.com>]

On Wed, Aug 11, 2021, Chenyi Qiang wrote:
> +	u32           pkrs;

...

> @@ -1115,6 +1117,7 @@ void vmx_prepare_switch_to_guest(struct kvm_vcpu *vcpu)
>  #endif
>  	unsigned long fs_base, gs_base;
>  	u16 fs_sel, gs_sel;
> +	u32 host_pkrs;

As mentioned in the previosu patch, I think it makes sense to track this as a u64
so that the only place in KVM that deas with the u64<=>u32 conversion is the below

	host_pkrs = get_current_pkrs();

>  	int i;
>  
>  	vmx->req_immediate_exit = false;
> @@ -1150,6 +1153,20 @@ void vmx_prepare_switch_to_guest(struct kvm_vcpu *vcpu)
>  	 */
>  	host_state->ldt_sel = kvm_read_ldt();
>  
> +	/*
> +	 * Update the host pkrs vmcs field before vcpu runs.
> +	 * The setting of VM_EXIT_LOAD_IA32_PKRS can ensure
> +	 * kvm_cpu_cap_has(X86_FEATURE_PKS) &&
> +	 * guest_cpuid_has(vcpu, X86_FEATURE_PKS)
> +	 */
> +	if (vm_exit_controls_get(vmx) & VM_EXIT_LOAD_IA32_PKRS) {
> +		host_pkrs = get_current_pkrs();
> +		if (unlikely(host_pkrs != host_state->pkrs)) {
> +			vmcs_write64(HOST_IA32_PKRS, host_pkrs);
> +			host_state->pkrs = host_pkrs;
> +		}
> +	}
> +
>  #ifdef CONFIG_X86_64
>  	savesegment(ds, host_state->ds_sel);
>  	savesegment(es, host_state->es_sel);
> @@ -1371,6 +1388,15 @@ void vmx_set_rflags(struct kvm_vcpu *vcpu, unsigned long rflags)
>  		vmx->emulation_required = emulation_required(vcpu);
>  }
>  
> +static void vmx_set_pkrs(struct kvm_vcpu *vcpu, u64 pkrs)
> +{

Hrm.  Ideally this would be open coded in vmx_set_msr().  Long term, the RESET/INIT
paths should really treat MSR updates as "normal" host_initiated writes instead of
having to manually handle every MSR.

That would be a bit gross to handle in vmx_vcpu_reset() since it would have to
create a struct msr_data (because __kvm_set_msr() isn't exposed to vendor code),
but since vcpu->arch.pkrs is relevant to the MMU I think it makes sense to
initiate the write from common x86.

E.g. this way there's not out-of-band special code, vmx_vcpu_reset() is kept clean,
and if/when SVM gains support for PKRS this particular path Just Works.  And it would
be an easy conversion for my pipe dream plan of handling MSRs at RESET/INIT via a
list of MSRs+values.

diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index ac83d873d65b..55881d13620f 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -11147,6 +11147,9 @@ void kvm_vcpu_reset(struct kvm_vcpu *vcpu, bool init_event)
        kvm_set_rflags(vcpu, X86_EFLAGS_FIXED);
        kvm_rip_write(vcpu, 0xfff0);

+       if (kvm_cpu_cap_has(X86_FEATURE_PKS))
+               __kvm_set_msr(vcpu, MSR_IA32_PKRS, 0, true);
+
        vcpu->arch.cr3 = 0;
        kvm_register_mark_dirty(vcpu, VCPU_EXREG_CR3);

> +	if (kvm_cpu_cap_has(X86_FEATURE_PKS)) {
> +		vcpu->arch.pkrs = pkrs;
> +		kvm_register_mark_available(vcpu, VCPU_EXREG_PKRS);
> +		vmcs_write64(GUEST_IA32_PKRS, pkrs);
> +	}
> +}