KVM: nVMX: Rework event injection and recovery
authorJan Kiszka <jan.kiszka@siemens.com>
Sun, 14 Apr 2013 10:12:46 +0000 (12:12 +0200)
committerGleb Natapov <gleb@redhat.com>
Sun, 14 Apr 2013 15:27:07 +0000 (18:27 +0300)
The basic idea is to always transfer the pending event injection on
vmexit into the architectural state of the VCPU and then drop it from
there if it turns out that we left L2 to enter L1, i.e. if we enter
prepare_vmcs12.

vmcs12_save_pending_events takes care to transfer pending L0 events into
the queue of L1. That is mandatory as L1 may decide to switch the guest
state completely, invalidating or preserving the pending events for
later injection (including on a different node, once we support
migration).

This concept is based on the rule that a pending vmlaunch/vmresume is
not canceled. Otherwise, we would risk to lose injected events or leak
them into the wrong queues. Encode this rule via a WARN_ON_ONCE at the
entry of nested_vmx_vmexit.

Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
arch/x86/kvm/vmx.c

index 5ae969b741f83bcc2dd0c802a97731cd21d81b06..4fb72a764dbd6a4b69c5ff48e80cdf211b9537c0 100644 (file)
@@ -6493,8 +6493,6 @@ static void __vmx_complete_interrupts(struct kvm_vcpu *vcpu,
 
 static void vmx_complete_interrupts(struct vcpu_vmx *vmx)
 {
-       if (is_guest_mode(&vmx->vcpu))
-               return;
        __vmx_complete_interrupts(&vmx->vcpu, vmx->idt_vectoring_info,
                                  VM_EXIT_INSTRUCTION_LEN,
                                  IDT_VECTORING_ERROR_CODE);
@@ -6502,8 +6500,6 @@ static void vmx_complete_interrupts(struct vcpu_vmx *vmx)
 
 static void vmx_cancel_injection(struct kvm_vcpu *vcpu)
 {
-       if (is_guest_mode(vcpu))
-               return;
        __vmx_complete_interrupts(vcpu,
                                  vmcs_read32(VM_ENTRY_INTR_INFO_FIELD),
                                  VM_ENTRY_INSTRUCTION_LEN,
@@ -6535,21 +6531,6 @@ static void __noclone vmx_vcpu_run(struct kvm_vcpu *vcpu)
        struct vcpu_vmx *vmx = to_vmx(vcpu);
        unsigned long debugctlmsr;
 
-       if (is_guest_mode(vcpu) && !vmx->nested.nested_run_pending) {
-               struct vmcs12 *vmcs12 = get_vmcs12(vcpu);
-               if (vmcs12->idt_vectoring_info_field &
-                               VECTORING_INFO_VALID_MASK) {
-                       vmcs_write32(VM_ENTRY_INTR_INFO_FIELD,
-                               vmcs12->idt_vectoring_info_field);
-                       vmcs_write32(VM_ENTRY_INSTRUCTION_LEN,
-                               vmcs12->vm_exit_instruction_len);
-                       if (vmcs12->idt_vectoring_info_field &
-                                       VECTORING_INFO_DELIVER_CODE_MASK)
-                               vmcs_write32(VM_ENTRY_EXCEPTION_ERROR_CODE,
-                                       vmcs12->idt_vectoring_error_code);
-               }
-       }
-
        /* Record the guest's net vcpu time for enforced NMI injections. */
        if (unlikely(!cpu_has_virtual_nmis() && vmx->soft_vnmi_blocked))
                vmx->entry_time = ktime_get();
@@ -6708,17 +6689,6 @@ static void __noclone vmx_vcpu_run(struct kvm_vcpu *vcpu)
 
        vmx->idt_vectoring_info = vmcs_read32(IDT_VECTORING_INFO_FIELD);
 
-       if (is_guest_mode(vcpu)) {
-               struct vmcs12 *vmcs12 = get_vmcs12(vcpu);
-               vmcs12->idt_vectoring_info_field = vmx->idt_vectoring_info;
-               if (vmx->idt_vectoring_info & VECTORING_INFO_VALID_MASK) {
-                       vmcs12->idt_vectoring_error_code =
-                               vmcs_read32(IDT_VECTORING_ERROR_CODE);
-                       vmcs12->vm_exit_instruction_len =
-                               vmcs_read32(VM_EXIT_INSTRUCTION_LEN);
-               }
-       }
-
        vmx->loaded_vmcs->launched = 1;
 
        vmx->exit_reason = vmcs_read32(VM_EXIT_REASON);
@@ -7326,6 +7296,48 @@ vmcs12_guest_cr4(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12)
                        vcpu->arch.cr4_guest_owned_bits));
 }
 
+static void vmcs12_save_pending_event(struct kvm_vcpu *vcpu,
+                                      struct vmcs12 *vmcs12)
+{
+       u32 idt_vectoring;
+       unsigned int nr;
+
+       if (vcpu->arch.exception.pending) {
+               nr = vcpu->arch.exception.nr;
+               idt_vectoring = nr | VECTORING_INFO_VALID_MASK;
+
+               if (kvm_exception_is_soft(nr)) {
+                       vmcs12->vm_exit_instruction_len =
+                               vcpu->arch.event_exit_inst_len;
+                       idt_vectoring |= INTR_TYPE_SOFT_EXCEPTION;
+               } else
+                       idt_vectoring |= INTR_TYPE_HARD_EXCEPTION;
+
+               if (vcpu->arch.exception.has_error_code) {
+                       idt_vectoring |= VECTORING_INFO_DELIVER_CODE_MASK;
+                       vmcs12->idt_vectoring_error_code =
+                               vcpu->arch.exception.error_code;
+               }
+
+               vmcs12->idt_vectoring_info_field = idt_vectoring;
+       } else if (vcpu->arch.nmi_pending) {
+               vmcs12->idt_vectoring_info_field =
+                       INTR_TYPE_NMI_INTR | INTR_INFO_VALID_MASK | NMI_VECTOR;
+       } else if (vcpu->arch.interrupt.pending) {
+               nr = vcpu->arch.interrupt.nr;
+               idt_vectoring = nr | VECTORING_INFO_VALID_MASK;
+
+               if (vcpu->arch.interrupt.soft) {
+                       idt_vectoring |= INTR_TYPE_SOFT_INTR;
+                       vmcs12->vm_entry_instruction_len =
+                               vcpu->arch.event_exit_inst_len;
+               } else
+                       idt_vectoring |= INTR_TYPE_EXT_INTR;
+
+               vmcs12->idt_vectoring_info_field = idt_vectoring;
+       }
+}
+
 /*
  * prepare_vmcs12 is part of what we need to do when the nested L2 guest exits
  * and we want to prepare to run its L1 parent. L1 keeps a vmcs for L2 (vmcs12),
@@ -7411,15 +7423,29 @@ static void prepare_vmcs12(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12)
 
        vmcs12->vm_exit_intr_info = vmcs_read32(VM_EXIT_INTR_INFO);
        vmcs12->vm_exit_intr_error_code = vmcs_read32(VM_EXIT_INTR_ERROR_CODE);
-       vmcs12->idt_vectoring_info_field = to_vmx(vcpu)->idt_vectoring_info;
-       vmcs12->idt_vectoring_error_code =
-               vmcs_read32(IDT_VECTORING_ERROR_CODE);
+       vmcs12->idt_vectoring_info_field = 0;
        vmcs12->vm_exit_instruction_len = vmcs_read32(VM_EXIT_INSTRUCTION_LEN);
        vmcs12->vmx_instruction_info = vmcs_read32(VMX_INSTRUCTION_INFO);
 
-       /* clear vm-entry fields which are to be cleared on exit */
-       if (!(vmcs12->vm_exit_reason & VMX_EXIT_REASONS_FAILED_VMENTRY))
+       if (!(vmcs12->vm_exit_reason & VMX_EXIT_REASONS_FAILED_VMENTRY)) {
+               /* vm_entry_intr_info_field is cleared on exit. Emulate this
+                * instead of reading the real value. */
                vmcs12->vm_entry_intr_info_field &= ~INTR_INFO_VALID_MASK;
+
+               /*
+                * Transfer the event that L0 or L1 may wanted to inject into
+                * L2 to IDT_VECTORING_INFO_FIELD.
+                */
+               vmcs12_save_pending_event(vcpu, vmcs12);
+       }
+
+       /*
+        * Drop what we picked up for L2 via vmx_complete_interrupts. It is
+        * preserved above and would only end up incorrectly in L1.
+        */
+       vcpu->arch.nmi_injected = false;
+       kvm_clear_exception_queue(vcpu);
+       kvm_clear_interrupt_queue(vcpu);
 }
 
 /*
@@ -7519,6 +7545,9 @@ static void nested_vmx_vmexit(struct kvm_vcpu *vcpu)
        int cpu;
        struct vmcs12 *vmcs12 = get_vmcs12(vcpu);
 
+       /* trying to cancel vmlaunch/vmresume is a bug */
+       WARN_ON_ONCE(vmx->nested.nested_run_pending);
+
        leave_guest_mode(vcpu);
        prepare_vmcs12(vcpu, vmcs12);