preempt_enable();
 }
 
-static bool start_hv_timer(struct kvm_lapic *apic)
+static bool __start_hv_timer(struct kvm_lapic *apic)
 {
-       u64 tscdeadline = apic->lapic_timer.tscdeadline;
+       struct kvm_timer *ktimer = &apic->lapic_timer;
+       int r;
 
-       if ((atomic_read(&apic->lapic_timer.pending) &&
-               !apic_lvtt_period(apic)) ||
-               kvm_x86_ops->set_hv_timer(apic->vcpu, tscdeadline)) {
-               if (apic->lapic_timer.hv_timer_in_use)
-                       cancel_hv_timer(apic);
-       } else {
-               apic->lapic_timer.hv_timer_in_use = true;
-               hrtimer_cancel(&apic->lapic_timer.timer);
+       if (!apic_lvtt_period(apic) && atomic_read(&ktimer->pending))
+               return false;
+
+       r = kvm_x86_ops->set_hv_timer(apic->vcpu, ktimer->tscdeadline);
+       if (r < 0)
+               return false;
+
+       ktimer->hv_timer_in_use = true;
+       hrtimer_cancel(&ktimer->timer);
 
-               /* In case the sw timer triggered in the window */
-               if (atomic_read(&apic->lapic_timer.pending) &&
-                       !apic_lvtt_period(apic))
+       /*
+        * Also recheck ktimer->pending, in case the sw timer triggered in
+        * the window.  For periodic timer, leave the hv timer running for
+        * simplicity, and the deadline will be recomputed on the next vmexit.
+        */
+       if (!apic_lvtt_period(apic) && atomic_read(&ktimer->pending))
+               return false;
+       return true;
+}
+
+static bool start_hv_timer(struct kvm_lapic *apic)
+{
+       if (!__start_hv_timer(apic)) {
+               if (apic->lapic_timer.hv_timer_in_use)
                        cancel_hv_timer(apic);
        }
+
        trace_kvm_hv_timer_state(apic->vcpu->vcpu_id,
                        apic->lapic_timer.hv_timer_in_use);
        return apic->lapic_timer.hv_timer_in_use;