Re: [PATCH] KVM: nVMX: set rflags to specify success in handle_invvpid() default case

[Vitaly Kuznetsov]

Sean Christopherson <sean.j.christopherson@xxxxxxxxx> writes:

> On Thu, Jan 23, 2020 at 10:22:24AM -0800, Jim Mattson wrote:
>> On Thu, Jan 23, 2020 at 1:54 AM Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote:
>> >
>> > On 23/01/20 10:45, Vitaly Kuznetsov wrote:
>> > >>> SDM says that "If an
>> > >>> unsupported INVVPID type is specified, the instruction fails." and this
>> > >>> is similar to INVEPT and I decided to check what handle_invept()
>> > >>> does. Well, it does BUG_ON().
>> > >>>
>> > >>> Are we doing the right thing in any of these cases?
>> > >>
>> > >> Yes, both INVEPT and INVVPID catch this earlier.
>> > >>
>> > >> So I'm leaning towards not applying Miaohe's patch.
>> > >
>> > > Well, we may at least want to converge on BUG_ON() for both
>> > > handle_invvpid()/handle_invept(), there's no need for them to differ.
>> >
>> > WARN_ON_ONCE + nested_vmx_failValid would probably be better, if we
>> > really want to change this.
>> >
>> > Paolo
>> 
>> In both cases, something is seriously wrong. The only plausible
>> explanations are compiler error or hardware failure. It would be nice
>> to handle *all* such failures with a KVM_INTERNAL_ERROR exit to
>> userspace. (I'm also thinking of situations like getting a VM-exit for
>> INIT.)
>
> Ya.  Vitaly and I had a similar discussion[*].  The idea we tossed around
> was to also mark the VM as having encountered a KVM/hardware bug so that
> the VM is effectively dead.  That would also allow gracefully handling bugs
> that are detected deep in the stack, i.e. can't simply return 0 to get out
> to userspace.

Yea, I was thinking about introducing a big hammer which would stop the
whole VM as soon as possible to make it easier to debug such
situations. Something like (not really tested):

diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index cf917139de6b..5476f88c9ada 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -8001,6 +8001,15 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
 	bool req_immediate_exit = false;
 
 	if (kvm_request_pending(vcpu)) {
+		/* INTERROR check should always come first */
+		if (kvm_check_request(KVM_REQ_INTERROR, vcpu)) {
+			if (vcpu->run->exit_reason != KVM_EXIT_INTERNAL_ERROR) {
+				vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
+				vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_OTHERCPU;
+			}
+			r = 0;
+			goto out;
+		}
 		if (kvm_check_request(KVM_REQ_GET_VMCS12_PAGES, vcpu)) {
 			if (unlikely(!kvm_x86_ops->get_vmcs12_pages(vcpu))) {
 				r = 0;
@@ -8510,6 +8519,13 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *kvm_run)
 	kvm_sigset_activate(vcpu);
 	kvm_load_guest_fpu(vcpu);
 
+	if (unlikely(vcpu->kvm->vm_bugged)) {
+		vcpu->run->exit_reason = KVM_REQ_INTERROR;
+		/* Maybe a suberror for 'attempted to run a vCPU of a bugged VM? */
+		r = 0;
+		goto out;
+	}
+
 	if (unlikely(vcpu->arch.mp_state == KVM_MP_STATE_UNINITIALIZED)) {
 		if (kvm_run->immediate_exit) {
 			r = -EINTR;
diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
index 538c25e778c0..d003be5fcf42 100644
--- a/include/linux/kvm_host.h
+++ b/include/linux/kvm_host.h
@@ -146,6 +146,7 @@ static inline bool is_error_page(struct page *page)
 #define KVM_REQ_MMU_RELOAD        (1 | KVM_REQUEST_WAIT | KVM_REQUEST_NO_WAKEUP)
 #define KVM_REQ_PENDING_TIMER     2
 #define KVM_REQ_UNHALT            3
+#define KVM_REQ_INTERROR          (4 | KVM_REQUEST_WAIT)
 #define KVM_REQUEST_ARCH_BASE     8
 
 #define KVM_ARCH_REQ_FLAGS(nr, flags) ({ \
@@ -501,6 +502,9 @@ struct kvm {
 	struct srcu_struct srcu;
 	struct srcu_struct irq_srcu;
 	pid_t userspace_pid;
+
+	/* VM caused internal KVM error */
+	bool vm_bugged;
 };
 
 #define kvm_err(fmt, ...) \
@@ -613,6 +617,7 @@ static inline void kvm_irqfd_exit(void)
 int kvm_init(void *opaque, unsigned vcpu_size, unsigned vcpu_align,
 		  struct module *module);
 void kvm_exit(void);
+void kvm_vm_bug(struct kvm_vcpu *vcpu, u32 error);
 
 void kvm_get_kvm(struct kvm *kvm);
 void kvm_put_kvm(struct kvm *kvm);
diff --git a/include/uapi/linux/kvm.h b/include/uapi/linux/kvm.h
index f0a16b4adbbd..62505161ae98 100644
--- a/include/uapi/linux/kvm.h
+++ b/include/uapi/linux/kvm.h
@@ -246,6 +246,8 @@ struct kvm_hyperv_exit {
 #define KVM_INTERNAL_ERROR_DELIVERY_EV	3
 /* Encounter unexpected vm-exit reason */
 #define KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON	4
+/* Some other vCPU caused internal KVM error */
+#define KVM_INTERNAL_ERROR_OTHERCPU	5
 
 /* for KVM_RUN, returned by mmap(vcpu_fd, offset=0) */
 struct kvm_run {
diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
index 00268290dcbd..4cc268d57714 100644
--- a/virt/kvm/kvm_main.c
+++ b/virt/kvm/kvm_main.c
@@ -4446,6 +4446,18 @@ void kvm_exit(void)
 }
 EXPORT_SYMBOL_GPL(kvm_exit);
 
+void kvm_vm_bug(struct kvm_vcpu *vcpu, u32 error)
+{
+	vcpu->kvm->vm_bugged = true;
+
+	vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
+	vcpu->run->internal.suberror = error;
+	/* We can also pass ndata/data ... */
+
+	kvm_make_all_cpus_request(vcpu->kvm, KVM_REQ_INTERROR);
+}
+EXPORT_SYMBOL_GPL(kvm_vm_bug);
+
 struct kvm_vm_worker_thread_context {
 	struct kvm *kvm;
 	struct task_struct *parent;

If you guys like the idea in general I can prepare patches.

-- 
Vitaly