KVM: SVM: Delay restoration of host MSR_TSC_AUX until return to userspace

Use KVM's "user return MSRs" framework to defer restoring the host's
MSR_TSC_AUX until the CPU returns to userspace.  Add/improve comments to
clarify why MSR_TSC_AUX is intercepted on both RDMSR and WRMSR, and why
it's safe for KVM to keep the guest's value loaded even if KVM is
scheduled out.

Cc: Reiji Watanabe <reijiw@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210423223404.3860547-5-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
index 9238831..3017925 100644 (file)
--- a/arch/x86/kvm/svm/svm.c
+++ b/arch/x86/kvm/svm/svm.c
@@ -213,6 +213,15 @@ struct kvm_ldttss_desc {
 
 DEFINE_PER_CPU(struct svm_cpu_data *, svm_data);
 
+/*
+ * Only MSR_TSC_AUX is switched via the user return hook.  EFER is switched via
+ * the VMCB, and the SYSCALL/SYSENTER MSRs are handled by VMLOAD/VMSAVE.
+ *
+ * RDTSCP and RDPID are not used in the kernel, specifically to allow KVM to
+ * defer the restoration of TSC_AUX until the CPU returns to userspace.
+ */
+#define TSC_AUX_URET_SLOT      0
+
 static const u32 msrpm_ranges[] = {0, 0xc0000000, 0xc0010000};
 
 #define NUM_MSR_MAPS ARRAY_SIZE(msrpm_ranges)
@@ -958,6 +967,9 @@ static __init int svm_hardware_setup(void)
                kvm_tsc_scaling_ratio_frac_bits = 32;
        }
 
+       if (boot_cpu_has(X86_FEATURE_RDTSCP))
+               kvm_define_user_return_msr(TSC_AUX_URET_SLOT, MSR_TSC_AUX);
+
        /* Check for pause filtering support */
        if (!boot_cpu_has(X86_FEATURE_PAUSEFILTER)) {
                pause_filter_count = 0;
@@ -1423,20 +1435,11 @@ static void svm_prepare_guest_switch(struct kvm_vcpu *vcpu)
 {
        struct vcpu_svm *svm = to_svm(vcpu);
        struct svm_cpu_data *sd = per_cpu(svm_data, vcpu->cpu);
-       unsigned int i;
 
        if (svm->guest_state_loaded)
                return;
 
        /*
-        * Certain MSRs are restored on VMEXIT (sev-es), or vmload of host save
-        * area (non-sev-es). Save ones that aren't so we can restore them
-        * individually later.
-        */
-       for (i = 0; i < NR_HOST_SAVE_USER_MSRS; i++)
-               rdmsrl(host_save_user_msrs[i], svm->host_user_msrs[i]);
-
-       /*
         * Save additional host state that will be restored on VMEXIT (sev-es)
         * or subsequent vmload of host save area.
         */
@@ -1454,29 +1457,15 @@ static void svm_prepare_guest_switch(struct kvm_vcpu *vcpu)
                }
        }
 
-       /* This assumes that the kernel never uses MSR_TSC_AUX */
        if (static_cpu_has(X86_FEATURE_RDTSCP))
-               wrmsrl(MSR_TSC_AUX, svm->tsc_aux);
+               kvm_set_user_return_msr(TSC_AUX_URET_SLOT, svm->tsc_aux, -1ull);
 
        svm->guest_state_loaded = true;
 }
 
 static void svm_prepare_host_switch(struct kvm_vcpu *vcpu)
 {
-       struct vcpu_svm *svm = to_svm(vcpu);
-       unsigned int i;
-
-       if (!svm->guest_state_loaded)
-               return;
-
-       /*
-        * Certain MSRs are restored on VMEXIT (sev-es), or vmload of host save
-        * area (non-sev-es). Restore the ones that weren't.
-        */
-       for (i = 0; i < NR_HOST_SAVE_USER_MSRS; i++)
-               wrmsrl(host_save_user_msrs[i], svm->host_user_msrs[i]);
-
-       svm->guest_state_loaded = false;
+       to_svm(vcpu)->guest_state_loaded = false;
 }
 
 static void svm_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
@@ -2790,6 +2779,7 @@ static int svm_set_vm_cr(struct kvm_vcpu *vcpu, u64 data)
 static int svm_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr)
 {
        struct vcpu_svm *svm = to_svm(vcpu);
+       int r;
 
        u32 ecx = msr->index;
        u64 data = msr->data;
@@ -2910,11 +2900,16 @@ static int svm_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr)
                data = (u32)data;
 
                /*
-                * This is rare, so we update the MSR here instead of using
-                * direct_access_msrs.  Doing that would require a rdmsr in
-                * svm_vcpu_put.
+                * TSC_AUX is usually changed only during boot and never read
+                * directly.  Intercept TSC_AUX instead of exposing it to the
+                * guest via direct_access_msrs, and switch it via user return.
                 */
-               wrmsrl(MSR_TSC_AUX, data);
+               preempt_disable();
+               r = kvm_set_user_return_msr(TSC_AUX_URET_SLOT, data, -1ull);
+               preempt_enable();
+               if (r)
+                       return 1;
+
                svm->tsc_aux = data;
                break;
        case MSR_IA32_DEBUGCTLMSR:

diff --git a/arch/x86/kvm/svm/svm.h b/arch/x86/kvm/svm/svm.h
index d620619..5d8027e 100644 (file)
--- a/arch/x86/kvm/svm/svm.h
+++ b/arch/x86/kvm/svm/svm.h
@@ -23,11 +23,6 @@
 
 #define __sme_page_pa(x) __sme_set(page_to_pfn(x) << PAGE_SHIFT)
 
-static const u32 host_save_user_msrs[] = {
-       MSR_TSC_AUX,
-};
-#define NR_HOST_SAVE_USER_MSRS ARRAY_SIZE(host_save_user_msrs)
-
 #define        IOPM_SIZE PAGE_SIZE * 3
 #define        MSRPM_SIZE PAGE_SIZE * 2
 
@@ -129,8 +124,6 @@ struct vcpu_svm {
 
        u64 next_rip;
 
-       u64 host_user_msrs[NR_HOST_SAVE_USER_MSRS];
-
        u64 spec_ctrl;
        /*
         * Contains guest-controlled bits of VIRT_SPEC_CTRL, which will be