KVM: LAPIC: add APIC Timer periodic/oneshot mode VMX preemption timer support
authorWanpeng Li <wanpeng.li@hotmail.com>
Mon, 24 Oct 2016 10:23:13 +0000 (18:23 +0800)
committerPaolo Bonzini <pbonzini@redhat.com>
Wed, 2 Nov 2016 20:32:17 +0000 (21:32 +0100)
Most windows guests still utilize APIC Timer periodic/oneshot mode
instead of tsc-deadline mode, and the APIC Timer periodic/oneshot
mode are still emulated by high overhead hrtimer on host. This patch
converts the expected expire time of the periodic/oneshot mode to
guest deadline tsc in order to leverage VMX preemption timer logic
for APIC Timer tsc-deadline mode. After each preemption timer vmexit
preemption timer is restarted to emulate LVTT current-count register
is automatically reloaded from the initial-count register when the
count reaches 0. This patch reduces ~5600 cycles for each APIC Timer
periodic mode operation virtualization.

Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Radim Krčmář <rkrcmar@redhat.com>
Cc: Yunhong Jiang <yunhong.jiang@intel.com>
Signed-off-by: Wanpeng Li <wanpeng.li@hotmail.com>
[Squashed with my fixes that were reviewed-by Paolo.]
Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>
arch/x86/kvm/lapic.c
arch/x86/kvm/lapic.h

index 0354a79fecd891839b11c8c4d6d47ff3b7f24b00..d1a2e60a1ef63d735ddbd705d32238e89d2e1223 100644 (file)
@@ -1090,7 +1090,7 @@ static void apic_send_ipi(struct kvm_lapic *apic)
 
 static u32 apic_get_tmcct(struct kvm_lapic *apic)
 {
-       ktime_t remaining;
+       ktime_t remaining, now;
        s64 ns;
        u32 tmcct;
 
@@ -1101,7 +1101,8 @@ static u32 apic_get_tmcct(struct kvm_lapic *apic)
                apic->lapic_timer.period == 0)
                return 0;
 
-       remaining = hrtimer_get_remaining(&apic->lapic_timer.timer);
+       now = apic->lapic_timer.timer.base->get_time();
+       remaining = ktime_sub(apic->lapic_timer.target_expiration, now);
        if (ktime_to_ns(remaining) < 0)
                remaining = ktime_set(0, 0);
 
@@ -1348,16 +1349,34 @@ static void start_sw_tscdeadline(struct kvm_lapic *apic)
 }
 
 static void start_sw_period(struct kvm_lapic *apic)
+{
+       if (!apic->lapic_timer.period)
+               return;
+
+       if (apic_lvtt_oneshot(apic) &&
+           ktime_after(apic->lapic_timer.timer.base->get_time(),
+                       apic->lapic_timer.target_expiration)) {
+               apic_timer_expired(apic);
+               return;
+       }
+
+       hrtimer_start(&apic->lapic_timer.timer,
+               apic->lapic_timer.target_expiration,
+               HRTIMER_MODE_ABS_PINNED);
+}
+
+static bool set_target_expiration(struct kvm_lapic *apic)
 {
        ktime_t now;
+       u64 tscl = rdtsc();
 
-       /* lapic timer in oneshot or periodic mode */
        now = apic->lapic_timer.timer.base->get_time();
        apic->lapic_timer.period = (u64)kvm_lapic_get_reg(apic, APIC_TMICT)
-                   * APIC_BUS_CYCLE_NS * apic->divide_count;
+               * APIC_BUS_CYCLE_NS * apic->divide_count;
 
        if (!apic->lapic_timer.period)
-               return;
+               return false;
+
        /*
         * Do not allow the guest to program periodic timers with small
         * interval, since the hrtimers are not throttled by the host
@@ -1376,10 +1395,6 @@ static void start_sw_period(struct kvm_lapic *apic)
                }
        }
 
-       hrtimer_start(&apic->lapic_timer.timer,
-                     ktime_add_ns(now, apic->lapic_timer.period),
-                     HRTIMER_MODE_ABS_PINNED);
-
        apic_debug("%s: bus cycle is %" PRId64 "ns, now 0x%016"
                   PRIx64 ", "
                   "timer initial count 0x%x, period %lldns, "
@@ -1389,6 +1404,21 @@ static void start_sw_period(struct kvm_lapic *apic)
                   apic->lapic_timer.period,
                   ktime_to_ns(ktime_add_ns(now,
                                apic->lapic_timer.period)));
+
+       apic->lapic_timer.tscdeadline = kvm_read_l1_tsc(apic->vcpu, tscl) +
+               nsec_to_cycles(apic->vcpu, apic->lapic_timer.period);
+       apic->lapic_timer.target_expiration = ktime_add_ns(now, apic->lapic_timer.period);
+
+       return true;
+}
+
+static void advance_periodic_target_expiration(struct kvm_lapic *apic)
+{
+       apic->lapic_timer.tscdeadline +=
+               nsec_to_cycles(apic->vcpu, apic->lapic_timer.period);
+       apic->lapic_timer.target_expiration =
+               ktime_add_ns(apic->lapic_timer.target_expiration,
+                               apic->lapic_timer.period);
 }
 
 bool kvm_lapic_hv_timer_in_use(struct kvm_vcpu *vcpu)
@@ -1406,22 +1436,12 @@ static void cancel_hv_timer(struct kvm_lapic *apic)
        apic->lapic_timer.hv_timer_in_use = false;
 }
 
-void kvm_lapic_expired_hv_timer(struct kvm_vcpu *vcpu)
-{
-       struct kvm_lapic *apic = vcpu->arch.apic;
-
-       WARN_ON(!apic->lapic_timer.hv_timer_in_use);
-       WARN_ON(swait_active(&vcpu->wq));
-       cancel_hv_timer(apic);
-       apic_timer_expired(apic);
-}
-EXPORT_SYMBOL_GPL(kvm_lapic_expired_hv_timer);
-
 static bool start_hv_timer(struct kvm_lapic *apic)
 {
        u64 tscdeadline = apic->lapic_timer.tscdeadline;
 
-       if (atomic_read(&apic->lapic_timer.pending) ||
+       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);
@@ -1430,7 +1450,8 @@ static bool start_hv_timer(struct kvm_lapic *apic)
                hrtimer_cancel(&apic->lapic_timer.timer);
 
                /* In case the sw timer triggered in the window */
-               if (atomic_read(&apic->lapic_timer.pending))
+               if (atomic_read(&apic->lapic_timer.pending) &&
+                       !apic_lvtt_period(apic))
                        cancel_hv_timer(apic);
        }
        trace_kvm_hv_timer_state(apic->vcpu->vcpu_id,
@@ -1438,14 +1459,30 @@ static bool start_hv_timer(struct kvm_lapic *apic)
        return apic->lapic_timer.hv_timer_in_use;
 }
 
+void kvm_lapic_expired_hv_timer(struct kvm_vcpu *vcpu)
+{
+       struct kvm_lapic *apic = vcpu->arch.apic;
+
+       WARN_ON(!apic->lapic_timer.hv_timer_in_use);
+       WARN_ON(swait_active(&vcpu->wq));
+       cancel_hv_timer(apic);
+       apic_timer_expired(apic);
+
+       if (apic_lvtt_period(apic) && apic->lapic_timer.period) {
+               advance_periodic_target_expiration(apic);
+               if (!start_hv_timer(apic))
+                       start_sw_period(apic);
+       }
+}
+EXPORT_SYMBOL_GPL(kvm_lapic_expired_hv_timer);
+
 void kvm_lapic_switch_to_hv_timer(struct kvm_vcpu *vcpu)
 {
        struct kvm_lapic *apic = vcpu->arch.apic;
 
        WARN_ON(apic->lapic_timer.hv_timer_in_use);
 
-       if (apic_lvtt_tscdeadline(apic))
-               start_hv_timer(apic);
+       start_hv_timer(apic);
 }
 EXPORT_SYMBOL_GPL(kvm_lapic_switch_to_hv_timer);
 
@@ -1462,7 +1499,10 @@ void kvm_lapic_switch_to_sw_timer(struct kvm_vcpu *vcpu)
        if (atomic_read(&apic->lapic_timer.pending))
                return;
 
-       start_sw_tscdeadline(apic);
+       if (apic_lvtt_period(apic) || apic_lvtt_oneshot(apic))
+               start_sw_period(apic);
+       else if (apic_lvtt_tscdeadline(apic))
+               start_sw_tscdeadline(apic);
 }
 EXPORT_SYMBOL_GPL(kvm_lapic_switch_to_sw_timer);
 
@@ -1470,9 +1510,11 @@ static void start_apic_timer(struct kvm_lapic *apic)
 {
        atomic_set(&apic->lapic_timer.pending, 0);
 
-       if (apic_lvtt_period(apic) || apic_lvtt_oneshot(apic))
-               start_sw_period(apic);
-       else if (apic_lvtt_tscdeadline(apic)) {
+       if (apic_lvtt_period(apic) || apic_lvtt_oneshot(apic)) {
+               if (set_target_expiration(apic) &&
+                       !(kvm_x86_ops->set_hv_timer && start_hv_timer(apic)))
+                       start_sw_period(apic);
+       } else if (apic_lvtt_tscdeadline(apic)) {
                if (!(kvm_x86_ops->set_hv_timer && start_hv_timer(apic)))
                        start_sw_tscdeadline(apic);
        }
@@ -1923,6 +1965,7 @@ static enum hrtimer_restart apic_timer_fn(struct hrtimer *data)
        apic_timer_expired(apic);
 
        if (lapic_is_periodic(apic)) {
+               advance_periodic_target_expiration(apic);
                hrtimer_add_expires_ns(&ktimer->timer, ktimer->period);
                return HRTIMER_RESTART;
        } else
@@ -2007,6 +2050,10 @@ void kvm_inject_apic_timer_irqs(struct kvm_vcpu *vcpu)
                kvm_apic_local_deliver(apic, APIC_LVTT);
                if (apic_lvtt_tscdeadline(apic))
                        apic->lapic_timer.tscdeadline = 0;
+               if (apic_lvtt_oneshot(apic)) {
+                       apic->lapic_timer.tscdeadline = 0;
+                       apic->lapic_timer.target_expiration = ktime_set(0, 0);
+               }
                atomic_set(&apic->lapic_timer.pending, 0);
        }
 }
index 031db2631a1a80142335a3d8bd4a6f6a01a2a248..e0c80233b3e17a1793cbafd422688091034b15c5 100644 (file)
@@ -15,6 +15,7 @@
 struct kvm_timer {
        struct hrtimer timer;
        s64 period;                             /* unit: ns */
+       ktime_t target_expiration;
        u32 timer_mode;
        u32 timer_mode_mask;
        u64 tscdeadline;