KVM: selftests: Add an x86-only test to verify nested exception queueing
authorSean Christopherson <seanjc@google.com>
Tue, 30 Aug 2022 23:16:13 +0000 (23:16 +0000)
committerPaolo Bonzini <pbonzini@redhat.com>
Mon, 26 Sep 2022 16:03:12 +0000 (12:03 -0400)
Add a test to verify that KVM_{G,S}ET_EVENTS play nice with pending vs.
injected exceptions when an exception is being queued for L2, and that
KVM correctly handles L1's exception intercept wants.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Link: https://lore.kernel.org/r/20220830231614.3580124-27-seanjc@google.com
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
tools/testing/selftests/kvm/.gitignore
tools/testing/selftests/kvm/Makefile
tools/testing/selftests/kvm/x86_64/nested_exceptions_test.c [new file with mode: 0644]

index d625a3f..45d9aee 100644 (file)
@@ -28,6 +28,7 @@
 /x86_64/max_vcpuid_cap_test
 /x86_64/mmio_warning_test
 /x86_64/monitor_mwait_test
+/x86_64/nested_exceptions_test
 /x86_64/nx_huge_pages_test
 /x86_64/platform_info_test
 /x86_64/pmu_event_filter_test
index 4c122f1..8b1b326 100644 (file)
@@ -89,6 +89,7 @@ TEST_GEN_PROGS_x86_64 += x86_64/kvm_clock_test
 TEST_GEN_PROGS_x86_64 += x86_64/kvm_pv_test
 TEST_GEN_PROGS_x86_64 += x86_64/mmio_warning_test
 TEST_GEN_PROGS_x86_64 += x86_64/monitor_mwait_test
+TEST_GEN_PROGS_x86_64 += x86_64/nested_exceptions_test
 TEST_GEN_PROGS_x86_64 += x86_64/platform_info_test
 TEST_GEN_PROGS_x86_64 += x86_64/pmu_event_filter_test
 TEST_GEN_PROGS_x86_64 += x86_64/set_boot_cpu_id
diff --git a/tools/testing/selftests/kvm/x86_64/nested_exceptions_test.c b/tools/testing/selftests/kvm/x86_64/nested_exceptions_test.c
new file mode 100644 (file)
index 0000000..ac33835
--- /dev/null
@@ -0,0 +1,295 @@
+// SPDX-License-Identifier: GPL-2.0-only
+#define _GNU_SOURCE /* for program_invocation_short_name */
+
+#include "test_util.h"
+#include "kvm_util.h"
+#include "processor.h"
+#include "vmx.h"
+#include "svm_util.h"
+
+#define L2_GUEST_STACK_SIZE 256
+
+/*
+ * Arbitrary, never shoved into KVM/hardware, just need to avoid conflict with
+ * the "real" exceptions used, #SS/#GP/#DF (12/13/8).
+ */
+#define FAKE_TRIPLE_FAULT_VECTOR       0xaa
+
+/* Arbitrary 32-bit error code injected by this test. */
+#define SS_ERROR_CODE 0xdeadbeef
+
+/*
+ * Bit '0' is set on Intel if the exception occurs while delivering a previous
+ * event/exception.  AMD's wording is ambiguous, but presumably the bit is set
+ * if the exception occurs while delivering an external event, e.g. NMI or INTR,
+ * but not for exceptions that occur when delivering other exceptions or
+ * software interrupts.
+ *
+ * Note, Intel's name for it, "External event", is misleading and much more
+ * aligned with AMD's behavior, but the SDM is quite clear on its behavior.
+ */
+#define ERROR_CODE_EXT_FLAG    BIT(0)
+
+/*
+ * Bit '1' is set if the fault occurred when looking up a descriptor in the
+ * IDT, which is the case here as the IDT is empty/NULL.
+ */
+#define ERROR_CODE_IDT_FLAG    BIT(1)
+
+/*
+ * The #GP that occurs when vectoring #SS should show the index into the IDT
+ * for #SS, plus have the "IDT flag" set.
+ */
+#define GP_ERROR_CODE_AMD ((SS_VECTOR * 8) | ERROR_CODE_IDT_FLAG)
+#define GP_ERROR_CODE_INTEL ((SS_VECTOR * 8) | ERROR_CODE_IDT_FLAG | ERROR_CODE_EXT_FLAG)
+
+/*
+ * Intel and AMD both shove '0' into the error code on #DF, regardless of what
+ * led to the double fault.
+ */
+#define DF_ERROR_CODE 0
+
+#define INTERCEPT_SS           (BIT_ULL(SS_VECTOR))
+#define INTERCEPT_SS_DF                (INTERCEPT_SS | BIT_ULL(DF_VECTOR))
+#define INTERCEPT_SS_GP_DF     (INTERCEPT_SS_DF | BIT_ULL(GP_VECTOR))
+
+static void l2_ss_pending_test(void)
+{
+       GUEST_SYNC(SS_VECTOR);
+}
+
+static void l2_ss_injected_gp_test(void)
+{
+       GUEST_SYNC(GP_VECTOR);
+}
+
+static void l2_ss_injected_df_test(void)
+{
+       GUEST_SYNC(DF_VECTOR);
+}
+
+static void l2_ss_injected_tf_test(void)
+{
+       GUEST_SYNC(FAKE_TRIPLE_FAULT_VECTOR);
+}
+
+static void svm_run_l2(struct svm_test_data *svm, void *l2_code, int vector,
+                      uint32_t error_code)
+{
+       struct vmcb *vmcb = svm->vmcb;
+       struct vmcb_control_area *ctrl = &vmcb->control;
+
+       vmcb->save.rip = (u64)l2_code;
+       run_guest(vmcb, svm->vmcb_gpa);
+
+       if (vector == FAKE_TRIPLE_FAULT_VECTOR)
+               return;
+
+       GUEST_ASSERT_EQ(ctrl->exit_code, (SVM_EXIT_EXCP_BASE + vector));
+       GUEST_ASSERT_EQ(ctrl->exit_info_1, error_code);
+}
+
+static void l1_svm_code(struct svm_test_data *svm)
+{
+       struct vmcb_control_area *ctrl = &svm->vmcb->control;
+       unsigned long l2_guest_stack[L2_GUEST_STACK_SIZE];
+
+       generic_svm_setup(svm, NULL, &l2_guest_stack[L2_GUEST_STACK_SIZE]);
+       svm->vmcb->save.idtr.limit = 0;
+       ctrl->intercept |= BIT_ULL(INTERCEPT_SHUTDOWN);
+
+       ctrl->intercept_exceptions = INTERCEPT_SS_GP_DF;
+       svm_run_l2(svm, l2_ss_pending_test, SS_VECTOR, SS_ERROR_CODE);
+       svm_run_l2(svm, l2_ss_injected_gp_test, GP_VECTOR, GP_ERROR_CODE_AMD);
+
+       ctrl->intercept_exceptions = INTERCEPT_SS_DF;
+       svm_run_l2(svm, l2_ss_injected_df_test, DF_VECTOR, DF_ERROR_CODE);
+
+       ctrl->intercept_exceptions = INTERCEPT_SS;
+       svm_run_l2(svm, l2_ss_injected_tf_test, FAKE_TRIPLE_FAULT_VECTOR, 0);
+       GUEST_ASSERT_EQ(ctrl->exit_code, SVM_EXIT_SHUTDOWN);
+
+       GUEST_DONE();
+}
+
+static void vmx_run_l2(void *l2_code, int vector, uint32_t error_code)
+{
+       GUEST_ASSERT(!vmwrite(GUEST_RIP, (u64)l2_code));
+
+       GUEST_ASSERT_EQ(vector == SS_VECTOR ? vmlaunch() : vmresume(), 0);
+
+       if (vector == FAKE_TRIPLE_FAULT_VECTOR)
+               return;
+
+       GUEST_ASSERT_EQ(vmreadz(VM_EXIT_REASON), EXIT_REASON_EXCEPTION_NMI);
+       GUEST_ASSERT_EQ((vmreadz(VM_EXIT_INTR_INFO) & 0xff), vector);
+       GUEST_ASSERT_EQ(vmreadz(VM_EXIT_INTR_ERROR_CODE), error_code);
+}
+
+static void l1_vmx_code(struct vmx_pages *vmx)
+{
+       unsigned long l2_guest_stack[L2_GUEST_STACK_SIZE];
+
+       GUEST_ASSERT_EQ(prepare_for_vmx_operation(vmx), true);
+
+       GUEST_ASSERT_EQ(load_vmcs(vmx), true);
+
+       prepare_vmcs(vmx, NULL, &l2_guest_stack[L2_GUEST_STACK_SIZE]);
+       GUEST_ASSERT_EQ(vmwrite(GUEST_IDTR_LIMIT, 0), 0);
+
+       /*
+        * VMX disallows injecting an exception with error_code[31:16] != 0,
+        * and hardware will never generate a VM-Exit with bits 31:16 set.
+        * KVM should likewise truncate the "bad" userspace value.
+        */
+       GUEST_ASSERT_EQ(vmwrite(EXCEPTION_BITMAP, INTERCEPT_SS_GP_DF), 0);
+       vmx_run_l2(l2_ss_pending_test, SS_VECTOR, (u16)SS_ERROR_CODE);
+       vmx_run_l2(l2_ss_injected_gp_test, GP_VECTOR, GP_ERROR_CODE_INTEL);
+
+       GUEST_ASSERT_EQ(vmwrite(EXCEPTION_BITMAP, INTERCEPT_SS_DF), 0);
+       vmx_run_l2(l2_ss_injected_df_test, DF_VECTOR, DF_ERROR_CODE);
+
+       GUEST_ASSERT_EQ(vmwrite(EXCEPTION_BITMAP, INTERCEPT_SS), 0);
+       vmx_run_l2(l2_ss_injected_tf_test, FAKE_TRIPLE_FAULT_VECTOR, 0);
+       GUEST_ASSERT_EQ(vmreadz(VM_EXIT_REASON), EXIT_REASON_TRIPLE_FAULT);
+
+       GUEST_DONE();
+}
+
+static void __attribute__((__flatten__)) l1_guest_code(void *test_data)
+{
+       if (this_cpu_has(X86_FEATURE_SVM))
+               l1_svm_code(test_data);
+       else
+               l1_vmx_code(test_data);
+}
+
+static void assert_ucall_vector(struct kvm_vcpu *vcpu, int vector)
+{
+       struct kvm_run *run = vcpu->run;
+       struct ucall uc;
+
+       TEST_ASSERT(run->exit_reason == KVM_EXIT_IO,
+                   "Unexpected exit reason: %u (%s),\n",
+                   run->exit_reason, exit_reason_str(run->exit_reason));
+
+       switch (get_ucall(vcpu, &uc)) {
+       case UCALL_SYNC:
+               TEST_ASSERT(vector == uc.args[1],
+                           "Expected L2 to ask for %d, got %ld", vector, uc.args[1]);
+               break;
+       case UCALL_DONE:
+               TEST_ASSERT(vector == -1,
+                           "Expected L2 to ask for %d, L2 says it's done", vector);
+               break;
+       case UCALL_ABORT:
+               TEST_FAIL("%s at %s:%ld (0x%lx != 0x%lx)",
+                         (const char *)uc.args[0], __FILE__, uc.args[1],
+                         uc.args[2], uc.args[3]);
+               break;
+       default:
+               TEST_FAIL("Expected L2 to ask for %d, got unexpected ucall %lu", vector, uc.cmd);
+       }
+}
+
+static void queue_ss_exception(struct kvm_vcpu *vcpu, bool inject)
+{
+       struct kvm_vcpu_events events;
+
+       vcpu_events_get(vcpu, &events);
+
+       TEST_ASSERT(!events.exception.pending,
+                   "Vector %d unexpectedlt pending", events.exception.nr);
+       TEST_ASSERT(!events.exception.injected,
+                   "Vector %d unexpectedly injected", events.exception.nr);
+
+       events.flags = KVM_VCPUEVENT_VALID_PAYLOAD;
+       events.exception.pending = !inject;
+       events.exception.injected = inject;
+       events.exception.nr = SS_VECTOR;
+       events.exception.has_error_code = true;
+       events.exception.error_code = SS_ERROR_CODE;
+       vcpu_events_set(vcpu, &events);
+}
+
+/*
+ * Verify KVM_{G,S}ET_EVENTS play nice with pending vs. injected exceptions
+ * when an exception is being queued for L2.  Specifically, verify that KVM
+ * honors L1 exception intercept controls when a #SS is pending/injected,
+ * triggers a #GP on vectoring the #SS, morphs to #DF if #GP isn't intercepted
+ * by L1, and finally causes (nested) SHUTDOWN if #DF isn't intercepted by L1.
+ */
+int main(int argc, char *argv[])
+{
+       vm_vaddr_t nested_test_data_gva;
+       struct kvm_vcpu_events events;
+       struct kvm_vcpu *vcpu;
+       struct kvm_vm *vm;
+
+       TEST_REQUIRE(kvm_has_cap(KVM_CAP_EXCEPTION_PAYLOAD));
+       TEST_REQUIRE(kvm_cpu_has(X86_FEATURE_SVM) || kvm_cpu_has(X86_FEATURE_VMX));
+
+       vm = vm_create_with_one_vcpu(&vcpu, l1_guest_code);
+       vm_enable_cap(vm, KVM_CAP_EXCEPTION_PAYLOAD, -2ul);
+
+       if (kvm_cpu_has(X86_FEATURE_SVM))
+               vcpu_alloc_svm(vm, &nested_test_data_gva);
+       else
+               vcpu_alloc_vmx(vm, &nested_test_data_gva);
+
+       vcpu_args_set(vcpu, 1, nested_test_data_gva);
+
+       /* Run L1 => L2.  L2 should sync and request #SS. */
+       vcpu_run(vcpu);
+       assert_ucall_vector(vcpu, SS_VECTOR);
+
+       /* Pend #SS and request immediate exit.  #SS should still be pending. */
+       queue_ss_exception(vcpu, false);
+       vcpu->run->immediate_exit = true;
+       vcpu_run_complete_io(vcpu);
+
+       /* Verify the pending events comes back out the same as it went in. */
+       vcpu_events_get(vcpu, &events);
+       ASSERT_EQ(events.flags & KVM_VCPUEVENT_VALID_PAYLOAD,
+                 KVM_VCPUEVENT_VALID_PAYLOAD);
+       ASSERT_EQ(events.exception.pending, true);
+       ASSERT_EQ(events.exception.nr, SS_VECTOR);
+       ASSERT_EQ(events.exception.has_error_code, true);
+       ASSERT_EQ(events.exception.error_code, SS_ERROR_CODE);
+
+       /*
+        * Run for real with the pending #SS, L1 should get a VM-Exit due to
+        * #SS interception and re-enter L2 to request #GP (via injected #SS).
+        */
+       vcpu->run->immediate_exit = false;
+       vcpu_run(vcpu);
+       assert_ucall_vector(vcpu, GP_VECTOR);
+
+       /*
+        * Inject #SS, the #SS should bypass interception and cause #GP, which
+        * L1 should intercept before KVM morphs it to #DF.  L1 should then
+        * disable #GP interception and run L2 to request #DF (via #SS => #GP).
+        */
+       queue_ss_exception(vcpu, true);
+       vcpu_run(vcpu);
+       assert_ucall_vector(vcpu, DF_VECTOR);
+
+       /*
+        * Inject #SS, the #SS should bypass interception and cause #GP, which
+        * L1 is no longer interception, and so should see a #DF VM-Exit.  L1
+        * should then signal that is done.
+        */
+       queue_ss_exception(vcpu, true);
+       vcpu_run(vcpu);
+       assert_ucall_vector(vcpu, FAKE_TRIPLE_FAULT_VECTOR);
+
+       /*
+        * Inject #SS yet again.  L1 is not intercepting #GP or #DF, and so
+        * should see nested TRIPLE_FAULT / SHUTDOWN.
+        */
+       queue_ss_exception(vcpu, true);
+       vcpu_run(vcpu);
+       assert_ucall_vector(vcpu, -1);
+
+       kvm_vm_free(vm);
+}