KVM: i8254: remove notifiers from PIT discard policy

Discard policy doesn't rely on information from notifiers, so we don't
need to register notifiers unconditionally.  We kept correct counts in
case userspace switched between policies during runtime, but that can be
avoided by reseting the state.
Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

arch/x86/kvm/i8254.c

@@ -225,7 +225,7 @@ static void kvm_pit_ack_irq(struct kvm_irq_ack_notifier *kian)
 	 * inc(pending) in pit_timer_fn and xchg(irq_ack, 0) in pit_do_work.
 	 */
 	smp_mb();
-	if (atomic_dec_if_positive(&ps->pending) > 0 && ps->reinject)
+	if (atomic_dec_if_positive(&ps->pending) > 0)
 		queue_kthread_work(&ps->pit->worker, &ps->pit->expired);
 }
 
@@ -301,6 +301,27 @@ static inline void kvm_pit_reset_reinject(struct kvm_pit *pit)
 	atomic_set(&pit->pit_state.irq_ack, 1);
 }
 
+void kvm_pit_set_reinject(struct kvm_pit *pit, bool reinject)
+{
+	struct kvm_kpit_state *ps = &pit->pit_state;
+	struct kvm *kvm = pit->kvm;
+
+	if (ps->reinject == reinject)
+		return;
+
+	if (reinject) {
+		/* The initial state is preserved while ps->reinject == 0. */
+		kvm_pit_reset_reinject(pit);
+		kvm_register_irq_ack_notifier(kvm, &ps->irq_ack_notifier);
+		kvm_register_irq_mask_notifier(kvm, 0, &pit->mask_notifier);
+	} else {
+		kvm_unregister_irq_ack_notifier(kvm, &ps->irq_ack_notifier);
+		kvm_unregister_irq_mask_notifier(kvm, 0, &pit->mask_notifier);
+	}
+
+	ps->reinject = reinject;
+}
+
 static void create_pit_timer(struct kvm_pit *pit, u32 val, int is_period)
 {
 	struct kvm_kpit_state *ps = &pit->pit_state;
@@ -681,15 +702,14 @@ struct kvm_pit *kvm_create_pit(struct kvm *kvm, u32 flags)
 	pit_state = &pit->pit_state;
 	pit_state->pit = pit;
 	hrtimer_init(&pit_state->timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
+
 	pit_state->irq_ack_notifier.gsi = 0;
 	pit_state->irq_ack_notifier.irq_acked = kvm_pit_ack_irq;
-	kvm_register_irq_ack_notifier(kvm, &pit_state->irq_ack_notifier);
-	pit_state->reinject = true;
+	pit->mask_notifier.func = pit_mask_notifer;
 
 	kvm_pit_reset(pit);
 
-	pit->mask_notifier.func = pit_mask_notifer;
-	kvm_register_irq_mask_notifier(kvm, 0, &pit->mask_notifier);
+	kvm_pit_set_reinject(pit, true);
 
 	kvm_iodevice_init(&pit->dev, &pit_dev_ops);
 	ret = kvm_io_bus_register_dev(kvm, KVM_PIO_BUS, KVM_PIT_BASE_ADDRESS,
@@ -712,8 +732,7 @@ struct kvm_pit *kvm_create_pit(struct kvm *kvm, u32 flags)
 	kvm_io_bus_unregister_dev(kvm, KVM_PIO_BUS, &pit->dev);
 
 fail:
-	kvm_unregister_irq_mask_notifier(kvm, 0, &pit->mask_notifier);
-	kvm_unregister_irq_ack_notifier(kvm, &pit_state->irq_ack_notifier);
+	kvm_pit_set_reinject(pit, false);
 	kvm_free_irq_source_id(kvm, pit->irq_source_id);
 	kthread_stop(pit->worker_task);
 	kfree(pit);
@@ -728,10 +747,7 @@ void kvm_free_pit(struct kvm *kvm)
 		kvm_io_bus_unregister_dev(kvm, KVM_PIO_BUS, &kvm->arch.vpit->dev);
 		kvm_io_bus_unregister_dev(kvm, KVM_PIO_BUS,
 					      &kvm->arch.vpit->speaker_dev);
-		kvm_unregister_irq_mask_notifier(kvm, 0,
-					       &kvm->arch.vpit->mask_notifier);
-		kvm_unregister_irq_ack_notifier(kvm,
-				&kvm->arch.vpit->pit_state.irq_ack_notifier);
+		kvm_pit_set_reinject(kvm->arch.vpit, false);
 		timer = &kvm->arch.vpit->pit_state.timer;
 		hrtimer_cancel(timer);
 		flush_kthread_work(&kvm->arch.vpit->expired);

arch/x86/kvm/i8254.h

@@ -63,5 +63,6 @@ void kvm_free_pit(struct kvm *kvm);
 
 void kvm_pit_load_count(struct kvm_pit *pit, int channel, u32 val,
 		int hpet_legacy_start);
+void kvm_pit_set_reinject(struct kvm_pit *pit, bool reinject);
 
 #endif

arch/x86/kvm/x86.c

@@ -3659,10 +3659,18 @@ static int kvm_vm_ioctl_set_pit2(struct kvm *kvm, struct kvm_pit_state2 *ps)
 static int kvm_vm_ioctl_reinject(struct kvm *kvm,
 				 struct kvm_reinject_control *control)
 {
-	if (!kvm->arch.vpit)
+	struct kvm_pit *pit = kvm->arch.vpit;
+
+	if (!pit)
 		return -ENXIO;
 
-	kvm->arch.vpit->pit_state.reinject = control->pit_reinject;
+	/* pit->pit_state.lock was overloaded to prevent userspace from getting
+	 * an inconsistent state after running multiple KVM_REINJECT_CONTROL
+	 * ioctls in parallel.  Use a separate lock if that ioctl isn't rare.
+	 */
+	mutex_lock(&pit->pit_state.lock);
+	kvm_pit_set_reinject(pit, control->pit_reinject);
+	mutex_unlock(&pit->pit_state.lock);
 
 	return 0;
 }