x86/kvm/mmu: introduce guest_mmu
authorVitaly Kuznetsov <vkuznets@redhat.com>
Mon, 8 Oct 2018 19:28:08 +0000 (21:28 +0200)
committerPaolo Bonzini <pbonzini@redhat.com>
Tue, 16 Oct 2018 22:30:04 +0000 (00:30 +0200)
When EPT is used for nested guest we need to re-init MMU as shadow
EPT MMU (nested_ept_init_mmu_context() does that). When we return back
from L2 to L1 kvm_mmu_reset_context() in nested_vmx_load_cr3() resets
MMU back to normal TDP mode. Add a special 'guest_mmu' so we can use
separate root caches; the improved hit rate is not very important for
single vCPU performance, but it avoids contention on the mmu_lock for
many vCPUs.

On the nested CPUID benchmark, with 16 vCPUs, an L2->L1->L2 vmexit
goes from 42k to 26k cycles.

Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Reviewed-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
arch/x86/include/asm/kvm_host.h
arch/x86/kvm/mmu.c
arch/x86/kvm/vmx.c

index 586ef14..7448d0b 100644 (file)
@@ -548,6 +548,9 @@ struct kvm_vcpu_arch {
        /* Non-nested MMU for L1 */
        struct kvm_mmu root_mmu;
 
+       /* L1 MMU when running nested */
+       struct kvm_mmu guest_mmu;
+
        /*
         * Paging state of an L2 guest (used for nested npt)
         *
index 96c7537..bf9e314 100644 (file)
@@ -4832,7 +4832,10 @@ EXPORT_SYMBOL_GPL(kvm_init_shadow_mmu);
 static union kvm_mmu_page_role
 kvm_calc_shadow_ept_root_page_role(struct kvm_vcpu *vcpu, bool accessed_dirty)
 {
-       union kvm_mmu_page_role role = vcpu->arch.mmu->base_role;
+       union kvm_mmu_page_role role;
+
+       /* Role is inherited from root_mmu */
+       role.word = vcpu->arch.root_mmu.base_role.word;
 
        role.level = PT64_ROOT_4LEVEL;
        role.direct = false;
@@ -4982,8 +4985,10 @@ EXPORT_SYMBOL_GPL(kvm_mmu_load);
 
 void kvm_mmu_unload(struct kvm_vcpu *vcpu)
 {
-       kvm_mmu_free_roots(vcpu, vcpu->arch.mmu, KVM_MMU_ROOTS_ALL);
-       WARN_ON(VALID_PAGE(vcpu->arch.mmu->root_hpa));
+       kvm_mmu_free_roots(vcpu, &vcpu->arch.root_mmu, KVM_MMU_ROOTS_ALL);
+       WARN_ON(VALID_PAGE(vcpu->arch.root_mmu.root_hpa));
+       kvm_mmu_free_roots(vcpu, &vcpu->arch.guest_mmu, KVM_MMU_ROOTS_ALL);
+       WARN_ON(VALID_PAGE(vcpu->arch.guest_mmu.root_hpa));
 }
 EXPORT_SYMBOL_GPL(kvm_mmu_unload);
 
@@ -5422,13 +5427,18 @@ int kvm_mmu_create(struct kvm_vcpu *vcpu)
 
        vcpu->arch.mmu = &vcpu->arch.root_mmu;
        vcpu->arch.walk_mmu = &vcpu->arch.root_mmu;
+
        vcpu->arch.root_mmu.root_hpa = INVALID_PAGE;
        vcpu->arch.root_mmu.translate_gpa = translate_gpa;
-       vcpu->arch.nested_mmu.translate_gpa = translate_nested_gpa;
-
        for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
                vcpu->arch.root_mmu.prev_roots[i] = KVM_MMU_ROOT_INFO_INVALID;
 
+       vcpu->arch.guest_mmu.root_hpa = INVALID_PAGE;
+       vcpu->arch.guest_mmu.translate_gpa = translate_gpa;
+       for (i = 0; i < KVM_MMU_NUM_PREV_ROOTS; i++)
+               vcpu->arch.guest_mmu.prev_roots[i] = KVM_MMU_ROOT_INFO_INVALID;
+
+       vcpu->arch.nested_mmu.translate_gpa = translate_nested_gpa;
        return alloc_mmu_pages(vcpu);
 }
 
index a0c3992..0288803 100644 (file)
@@ -8423,8 +8423,10 @@ static void vmx_disable_shadow_vmcs(struct vcpu_vmx *vmx)
        vmcs_write64(VMCS_LINK_POINTER, -1ull);
 }
 
-static inline void nested_release_vmcs12(struct vcpu_vmx *vmx)
+static inline void nested_release_vmcs12(struct kvm_vcpu *vcpu)
 {
+       struct vcpu_vmx *vmx = to_vmx(vcpu);
+
        if (vmx->nested.current_vmptr == -1ull)
                return;
 
@@ -8438,10 +8440,12 @@ static inline void nested_release_vmcs12(struct vcpu_vmx *vmx)
        vmx->nested.posted_intr_nv = -1;
 
        /* Flush VMCS12 to guest memory */
-       kvm_vcpu_write_guest_page(&vmx->vcpu,
+       kvm_vcpu_write_guest_page(vcpu,
                                  vmx->nested.current_vmptr >> PAGE_SHIFT,
                                  vmx->nested.cached_vmcs12, 0, VMCS12_SIZE);
 
+       kvm_mmu_free_roots(vcpu, &vcpu->arch.guest_mmu, KVM_MMU_ROOTS_ALL);
+
        vmx->nested.current_vmptr = -1ull;
 }
 
@@ -8449,8 +8453,10 @@ static inline void nested_release_vmcs12(struct vcpu_vmx *vmx)
  * Free whatever needs to be freed from vmx->nested when L1 goes down, or
  * just stops using VMX.
  */
-static void free_nested(struct vcpu_vmx *vmx)
+static void free_nested(struct kvm_vcpu *vcpu)
 {
+       struct vcpu_vmx *vmx = to_vmx(vcpu);
+
        if (!vmx->nested.vmxon && !vmx->nested.smm.vmxon)
                return;
 
@@ -8483,6 +8489,8 @@ static void free_nested(struct vcpu_vmx *vmx)
                vmx->nested.pi_desc = NULL;
        }
 
+       kvm_mmu_free_roots(vcpu, &vcpu->arch.guest_mmu, KVM_MMU_ROOTS_ALL);
+
        free_loaded_vmcs(&vmx->nested.vmcs02);
 }
 
@@ -8491,7 +8499,7 @@ static int handle_vmoff(struct kvm_vcpu *vcpu)
 {
        if (!nested_vmx_check_permission(vcpu))
                return 1;
-       free_nested(to_vmx(vcpu));
+       free_nested(vcpu);
        return nested_vmx_succeed(vcpu);
 }
 
@@ -8517,7 +8525,7 @@ static int handle_vmclear(struct kvm_vcpu *vcpu)
                        VMXERR_VMCLEAR_VMXON_POINTER);
 
        if (vmptr == vmx->nested.current_vmptr)
-               nested_release_vmcs12(vmx);
+               nested_release_vmcs12(vcpu);
 
        kvm_vcpu_write_guest(vcpu,
                        vmptr + offsetof(struct vmcs12, launch_state),
@@ -8872,7 +8880,8 @@ static int handle_vmptrld(struct kvm_vcpu *vcpu)
                                VMXERR_VMPTRLD_INCORRECT_VMCS_REVISION_ID);
                }
 
-               nested_release_vmcs12(vmx);
+               nested_release_vmcs12(vcpu);
+
                /*
                 * Load VMCS12 from guest memory since it is not already
                 * cached.
@@ -10928,12 +10937,10 @@ static void vmx_switch_vmcs(struct kvm_vcpu *vcpu, struct loaded_vmcs *vmcs)
  */
 static void vmx_free_vcpu_nested(struct kvm_vcpu *vcpu)
 {
-       struct vcpu_vmx *vmx = to_vmx(vcpu);
-
-       vcpu_load(vcpu);
-       vmx_switch_vmcs(vcpu, &vmx->vmcs01);
-       free_nested(vmx);
-       vcpu_put(vcpu);
+       vcpu_load(vcpu);
+       vmx_switch_vmcs(vcpu, &to_vmx(vcpu)->vmcs01);
+       free_nested(vcpu);
+       vcpu_put(vcpu);
 }
 
 static void vmx_free_vcpu(struct kvm_vcpu *vcpu)
@@ -11300,6 +11307,7 @@ static void nested_ept_init_mmu_context(struct kvm_vcpu *vcpu)
 {
        WARN_ON(mmu_is_nested(vcpu));
 
+       vcpu->arch.mmu = &vcpu->arch.guest_mmu;
        kvm_init_shadow_ept_mmu(vcpu,
                        to_vmx(vcpu)->nested.msrs.ept_caps &
                        VMX_EPT_EXECUTE_ONLY_BIT,
@@ -11315,6 +11323,7 @@ static void nested_ept_init_mmu_context(struct kvm_vcpu *vcpu)
 
 static void nested_ept_uninit_mmu_context(struct kvm_vcpu *vcpu)
 {
+       vcpu->arch.mmu = &vcpu->arch.root_mmu;
        vcpu->arch.walk_mmu = &vcpu->arch.root_mmu;
 }
 
@@ -13731,7 +13740,7 @@ static void vmx_leave_nested(struct kvm_vcpu *vcpu)
                to_vmx(vcpu)->nested.nested_run_pending = 0;
                nested_vmx_vmexit(vcpu, -1, 0, 0);
        }
-       free_nested(to_vmx(vcpu));
+       free_nested(vcpu);
 }
 
 static int vmx_check_intercept(struct kvm_vcpu *vcpu,