Re: [RFC 2/7] KVM: VMX: Expose IA32_PKRS MSR

From: Chenyi Qiang
Date: Thu Aug 13 2020 - 01:42:41 EST




On 8/13/2020 5:21 AM, Jim Mattson wrote:
On Fri, Aug 7, 2020 at 1:46 AM Chenyi Qiang <chenyi.qiang@xxxxxxxxx> wrote:

Protection Keys for Supervisor Pages (PKS) uses IA32_PKRS MSR (PKRS) at
index 0x6E1 to allow software to manage supervisor protection key
rights. For performance consideration, PKRS intercept will be disabled
so that the guest can access the PKRS without VM exits.
PKS introduces dedicated control fields in VMCS to switch PKRS, which
only does the retore part. In addition, every VM exit saves PKRS into
the guest-state area in VMCS, while VM enter won't save the host value
due to the expectation that the host won't change the MSR often. Update
the host's value in VMCS manually if the MSR has been changed by the
kernel since the last time the VMCS was run.
The function get_current_pkrs() in arch/x86/mm/pkeys.c exports the
per-cpu variable pkrs_cache to avoid frequent rdmsr of PKRS.

Signed-off-by: Chenyi Qiang <chenyi.qiang@xxxxxxxxx>
---

diff --git a/arch/x86/kvm/vmx/nested.c b/arch/x86/kvm/vmx/nested.c
index 11e4df560018..df2c2e733549 100644
--- a/arch/x86/kvm/vmx/nested.c
+++ b/arch/x86/kvm/vmx/nested.c
@@ -289,6 +289,7 @@ static void vmx_sync_vmcs_host_state(struct vcpu_vmx *vmx,
dest->ds_sel = src->ds_sel;
dest->es_sel = src->es_sel;
#endif
+ dest->pkrs = src->pkrs;

Why isn't this (and other PKRS code) inside the #ifdef CONFIG_X86_64?
PKRS isn't usable outside of long mode, is it?


Yes, I'm also thinking about whether to put all pks code into CONFIG_X86_64. The kernel implementation also wrap its pks code inside CONFIG_ARCH_HAS_SUPERVISOR_PKEYS which has dependency with CONFIG_X86_64.
However, maybe this can help when host kernel disable PKS but the guest enable it. What do you think about this?


}

static void vmx_switch_vmcs(struct kvm_vcpu *vcpu, struct loaded_vmcs *vmcs)
diff --git a/arch/x86/kvm/vmx/vmcs.h b/arch/x86/kvm/vmx/vmcs.h
index 7a3675fddec2..39ec3d0c844b 100644
--- a/arch/x86/kvm/vmx/vmcs.h
+++ b/arch/x86/kvm/vmx/vmcs.h
@@ -40,6 +40,7 @@ struct vmcs_host_state {
#ifdef CONFIG_X86_64
u16 ds_sel, es_sel;
#endif
+ u32 pkrs;

One thing that does seem odd throughout is that PKRS is a 64-bit
resource, but we store and pass around only 32-bits. Yes, the top 32
bits are currently reserved, but the same can be said of, say, cr4, a
few lines above this. Yet, we store and pass around cr4 as 64-bits.
How do we decide?


IMO, If the high part of PKRS is zero-reserved, it's OK to use u32. I define it as u32 just to follow the definition pkrs_cache in kernel code.

};

struct vmcs_controls_shadow {

@@ -1163,6 +1164,20 @@ 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_VMX).
+ */
+ 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);
@@ -1951,6 +1966,13 @@ static int vmx_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
else
msr_info->data = vmx->pt_desc.guest.addr_a[index / 2];
break;
+ case MSR_IA32_PKRS:
+ if (!kvm_cpu_cap_has(X86_FEATURE_PKS) ||
+ (!msr_info->host_initiated &&
+ !guest_cpuid_has(vcpu, X86_FEATURE_PKS)))

Could this be simplified if we required
kvm_cpu_cap_has(X86_FEATURE_PKS) as a prerequisite for
guest_cpuid_has(vcpu, X86_FEATURE_PKS)? If not, what is the expected
behavior if guest_cpuid_has(vcpu, X86_FEATURE_PKS) is true and
kvm_cpu_cap_has(X86_FEATURE_PKS) is false?


Yes, kvm_cpu_cap_has() is a prerequisite for guest_cpuid_has() as we have done the check in vmx_cpuid_update(). Here I add the kvm_cpu_cap_has() check to ensure the vmcs_read(GUEST_IA32_PKRS) can execute correctly in case of the userspace access.

+ return 1;
+ msr_info->data = vmcs_read64(GUEST_IA32_PKRS);
+ break;
case MSR_TSC_AUX:
if (!msr_info->host_initiated &&
!guest_cpuid_has(vcpu, X86_FEATURE_RDTSCP))

@@ -7230,6 +7267,26 @@ static void update_intel_pt_cfg(struct kvm_vcpu *vcpu)
vmx->pt_desc.ctl_bitmask &= ~(0xfULL << (32 + i * 4));
}

+static void vmx_update_pkrs_cfg(struct kvm_vcpu *vcpu)
+{
+ struct vcpu_vmx *vmx = to_vmx(vcpu);
+ unsigned long *msr_bitmap = vmx->vmcs01.msr_bitmap;
+ bool pks_supported = guest_cpuid_has(vcpu, X86_FEATURE_PKS);
+
+ /*
+ * set intercept for PKRS when the guest doesn't support pks
+ */
+ vmx_set_intercept_for_msr(msr_bitmap, MSR_IA32_PKRS, MSR_TYPE_RW, !pks_supported);
+
+ if (pks_supported) {
+ vm_entry_controls_setbit(vmx, VM_ENTRY_LOAD_IA32_PKRS);
+ vm_exit_controls_setbit(vmx, VM_EXIT_LOAD_IA32_PKRS);
+ } else {
+ vm_entry_controls_clearbit(vmx, VM_ENTRY_LOAD_IA32_PKRS);
+ vm_exit_controls_clearbit(vmx, VM_EXIT_LOAD_IA32_PKRS);
+ }
+}

Not just your change, but it is unclear to me what the expected
behavior is when CPUID bits are modified while the guest is running.
For example, it seems that this code results in immediate behavioral
changes for an L1 guest; however, if an L2 guest is active, then there
are no behavioral changes until the next emulated VM-entry from L1 to
L2. Is that right?


I don't know if there is a way to deal with the CPUID modification in KVM while the guest is running. Some CPUID modification like X86_FEATURE_OSPKE happens when the guest sets CR4_PKE. But I'm not familiar with your case.

Paolo

What's your opinion?


On a more general note, do you have any kvm-unit-tests that exercise
the new code?


Yes, I'll attach the kvm-unit-tests in next version.