KVM: nVMX: fixes to nested virt interrupt injection
authorPaolo Bonzini <pbonzini@redhat.com>
Thu, 27 Jul 2017 10:29:32 +0000 (12:29 +0200)
committerRadim Krčmář <rkrcmar@redhat.com>
Tue, 1 Aug 2017 20:24:17 +0000 (22:24 +0200)
There are three issues in nested_vmx_check_exception:

1) it is not taking PFEC_MATCH/PFEC_MASK into account, as reported
by Wanpeng Li;

2) it should rebuild the interruption info and exit qualification fields
from scratch, as reported by Jim Mattson, because the values from the
L2->L0 vmexit may be invalid (e.g. if an emulated instruction causes
a page fault, the EPT misconfig's exit qualification is incorrect).

3) CR2 and DR6 should not be written for exception intercept vmexits
(CR2 only for AMD).

This patch fixes the first two and adds a comment about the last,
outlining the fix.

Cc: Jim Mattson <jmattson@google.com>
Cc: Wanpeng Li <wanpeng.li@hotmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
arch/x86/kvm/svm.c
arch/x86/kvm/vmx.c

index 4d8141e533c369711df245d0a783683598ad4559..1107626938ccff4f26da130f8b570928ca7c8efb 100644 (file)
@@ -2430,6 +2430,16 @@ static int nested_svm_check_exception(struct vcpu_svm *svm, unsigned nr,
        svm->vmcb->control.exit_code = SVM_EXIT_EXCP_BASE + nr;
        svm->vmcb->control.exit_code_hi = 0;
        svm->vmcb->control.exit_info_1 = error_code;
+
+       /*
+        * FIXME: we should not write CR2 when L1 intercepts an L2 #PF exception.
+        * The fix is to add the ancillary datum (CR2 or DR6) to structs
+        * kvm_queued_exception and kvm_vcpu_events, so that CR2 and DR6 can be
+        * written only when inject_pending_event runs (DR6 would written here
+        * too).  This should be conditional on a new capability---if the
+        * capability is disabled, kvm_multiple_exception would write the
+        * ancillary information to CR2 or DR6, for backwards ABI-compatibility.
+        */
        if (svm->vcpu.arch.exception.nested_apf)
                svm->vmcb->control.exit_info_2 = svm->vcpu.arch.apf.nested_apf_token;
        else
index 19465b73cc07fbf2196ef05dbc084a0cd28e991c..714d4364ef87b0def49700dab228d70d573c7b10 100644 (file)
@@ -927,6 +927,10 @@ static u32 vmx_segment_access_rights(struct kvm_segment *var);
 static void copy_vmcs12_to_shadow(struct vcpu_vmx *vmx);
 static void copy_shadow_to_vmcs12(struct vcpu_vmx *vmx);
 static int alloc_identity_pagetable(struct kvm *kvm);
+static bool vmx_get_nmi_mask(struct kvm_vcpu *vcpu);
+static void vmx_set_nmi_mask(struct kvm_vcpu *vcpu, bool masked);
+static bool nested_vmx_is_page_fault_vmexit(struct vmcs12 *vmcs12,
+                                           u16 error_code);
 
 static DEFINE_PER_CPU(struct vmcs *, vmxarea);
 static DEFINE_PER_CPU(struct vmcs *, current_vmcs);
@@ -2428,6 +2432,30 @@ static void skip_emulated_instruction(struct kvm_vcpu *vcpu)
        vmx_set_interrupt_shadow(vcpu, 0);
 }
 
+static void nested_vmx_inject_exception_vmexit(struct kvm_vcpu *vcpu,
+                                              unsigned long exit_qual)
+{
+       struct vmcs12 *vmcs12 = get_vmcs12(vcpu);
+       unsigned int nr = vcpu->arch.exception.nr;
+       u32 intr_info = nr | INTR_INFO_VALID_MASK;
+
+       if (vcpu->arch.exception.has_error_code) {
+               vmcs12->vm_exit_intr_error_code = vcpu->arch.exception.error_code;
+               intr_info |= INTR_INFO_DELIVER_CODE_MASK;
+       }
+
+       if (kvm_exception_is_soft(nr))
+               intr_info |= INTR_TYPE_SOFT_EXCEPTION;
+       else
+               intr_info |= INTR_TYPE_HARD_EXCEPTION;
+
+       if (!(vmcs12->idt_vectoring_info_field & VECTORING_INFO_VALID_MASK) &&
+           vmx_get_nmi_mask(vcpu))
+               intr_info |= INTR_INFO_UNBLOCK_NMI;
+
+       nested_vmx_vmexit(vcpu, EXIT_REASON_EXCEPTION_NMI, intr_info, exit_qual);
+}
+
 /*
  * KVM wants to inject page-faults which it got to the guest. This function
  * checks whether in a nested guest, we need to inject them to L1 or L2.
@@ -2437,24 +2465,38 @@ static int nested_vmx_check_exception(struct kvm_vcpu *vcpu)
        struct vmcs12 *vmcs12 = get_vmcs12(vcpu);
        unsigned int nr = vcpu->arch.exception.nr;
 
-       if (!((vmcs12->exception_bitmap & (1u << nr)) ||
-               (nr == PF_VECTOR && vcpu->arch.exception.nested_apf)))
-               return 0;
+       if (nr == PF_VECTOR) {
+               if (vcpu->arch.exception.nested_apf) {
+                       nested_vmx_inject_exception_vmexit(vcpu,
+                                                          vcpu->arch.apf.nested_apf_token);
+                       return 1;
+               }
+               /*
+                * FIXME: we must not write CR2 when L1 intercepts an L2 #PF exception.
+                * The fix is to add the ancillary datum (CR2 or DR6) to structs
+                * kvm_queued_exception and kvm_vcpu_events, so that CR2 and DR6
+                * can be written only when inject_pending_event runs.  This should be
+                * conditional on a new capability---if the capability is disabled,
+                * kvm_multiple_exception would write the ancillary information to
+                * CR2 or DR6, for backwards ABI-compatibility.
+                */
+               if (nested_vmx_is_page_fault_vmexit(vmcs12,
+                                                   vcpu->arch.exception.error_code)) {
+                       nested_vmx_inject_exception_vmexit(vcpu, vcpu->arch.cr2);
+                       return 1;
+               }
+       } else {
+               unsigned long exit_qual = 0;
+               if (nr == DB_VECTOR)
+                       exit_qual = vcpu->arch.dr6;
 
-       if (vcpu->arch.exception.nested_apf) {
-               vmcs12->vm_exit_intr_error_code = vcpu->arch.exception.error_code;
-               nested_vmx_vmexit(vcpu, EXIT_REASON_EXCEPTION_NMI,
-                       PF_VECTOR | INTR_TYPE_HARD_EXCEPTION |
-                       INTR_INFO_DELIVER_CODE_MASK | INTR_INFO_VALID_MASK,
-                       vcpu->arch.apf.nested_apf_token);
-               return 1;
+               if (vmcs12->exception_bitmap & (1u << nr)) {
+                       nested_vmx_inject_exception_vmexit(vcpu, exit_qual);
+                       return 1;
+               }
        }
 
-       vmcs12->vm_exit_intr_error_code = vmcs_read32(VM_EXIT_INTR_ERROR_CODE);
-       nested_vmx_vmexit(vcpu, EXIT_REASON_EXCEPTION_NMI,
-                         vmcs_read32(VM_EXIT_INTR_INFO),
-                         vmcs_readl(EXIT_QUALIFICATION));
-       return 1;
+       return 0;
 }
 
 static void vmx_queue_exception(struct kvm_vcpu *vcpu)
@@ -9529,10 +9571,11 @@ static void vmx_inject_page_fault_nested(struct kvm_vcpu *vcpu,
        WARN_ON(!is_guest_mode(vcpu));
 
        if (nested_vmx_is_page_fault_vmexit(vmcs12, fault->error_code)) {
-               vmcs12->vm_exit_intr_error_code = vmcs_read32(VM_EXIT_INTR_ERROR_CODE);
-               nested_vmx_vmexit(vcpu, to_vmx(vcpu)->exit_reason,
-                                 vmcs_read32(VM_EXIT_INTR_INFO),
-                                 vmcs_readl(EXIT_QUALIFICATION));
+               vmcs12->vm_exit_intr_error_code = fault->error_code;
+               nested_vmx_vmexit(vcpu, EXIT_REASON_EXCEPTION_NMI,
+                                 PF_VECTOR | INTR_TYPE_HARD_EXCEPTION |
+                                 INTR_INFO_DELIVER_CODE_MASK | INTR_INFO_VALID_MASK,
+                                 fault->address);
        } else {
                kvm_inject_page_fault(vcpu, fault);
        }
@@ -10115,12 +10158,6 @@ static int prepare_vmcs02(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12,
         * "or"ing of the EB of vmcs01 and vmcs12, because when enable_ept,
         * vmcs01's EB.PF is 0 so the "or" will take vmcs12's value, and when
         * !enable_ept, EB.PF is 1, so the "or" will always be 1.
-        *
-        * A problem with this approach (when !enable_ept) is that L1 may be
-        * injected with more page faults than it asked for. This could have
-        * caused problems, but in practice existing hypervisors don't care.
-        * To fix this, we will need to emulate the PFEC checking (on the L1
-        * page tables), using walk_addr(), when injecting PFs to L1.
         */
        vmcs_write32(PAGE_FAULT_ERROR_CODE_MASK,
                enable_ept ? vmcs12->page_fault_error_code_mask : 0);