sched/idle: Prevent pointless NOHZ transitions in default_idle_call()

kernel/sched/idle.c +3 -2

+static inline void default_idle_enter(void) { }
+#endif /* !CONFIG_NO_HZ_COMMON */
+
 static int call_cpuidle_s2idle(struct cpuidle_driver *drv,
 			       struct cpuidle_device *dev,
 			       u64 max_latency_ns)

@@ -186,8 +236,6 @@ static void cpuidle_idle_call(void)
 	}
 
 	if (cpuidle_not_available(drv, dev)) {
-		tick_nohz_idle_stop_tick();
-
 		default_idle_call();
 		goto exit_idle;
 	}

kernel/sched/idle.c +0 -0

 	while (!need_resched()) {
 
 

Guests fall back to default_idle_call() as there is no cpuidle driver available to them by default. That causes a problem in fully loaded scenarios where CPUs go briefly idle for a couple of microseconds:

tick_nohz_idle_stop_tick() is invoked unconditionally which means unless there is timer pending in the next tick, the tick is stopped and a couple of microseconds later when the idle condition goes away restarted. That requires to program the clockevent device twice which implies a VM exit for each reprogramming.

It was suggested to remove the tick_nohz_idle_stop_tick() invocation from the default idle code, but would be counterproductive. It would not allow the host to go into deeper idle states when the guest CPU is fully idle as it has to maintain the periodic tick.

Cure this by implementing a trivial moving average filter which keeps track of the recent idle recidency time and only stop the tick when the average is larger than a tick.

Signed-off-by Thomas Gleixner <[email protected]>

Quoted 15 lines · Thomas Gleixner

Guests fall back to default_idle_call() as there is no cpuidle driver available to them by default. That causes a problem in fully loaded scenarios where CPUs go briefly idle for a couple of microseconds:

tick_nohz_idle_stop_tick() is invoked unconditionally which means unless there is timer pending in the next tick, the tick is stopped and a couple of microseconds later when the idle condition goes away restarted. That requires to program the clockevent device twice which implies a VM exit for each reprogramming.

It was suggested to remove the tick_nohz_idle_stop_tick() invocation from the default idle code, but would be counterproductive. It would not allow the host to go into deeper idle states when the guest CPU is fully idle as it has to maintain the periodic tick.

Quoted 5 lines

Cure this by implementing a trivial moving average filter which keeps track of the recent idle recidency time and only stop the tick when the average is larger than a tick.

Signed-off-by Thomas Gleixner <[email protected]>

Shouldn't there be instead a new dedicated cpuidle driver with proper governor support?

Quoted 20 lines · Frederic Weisbecker

Quoted 19 lines · Thomas Gleixner

Guests fall back to default_idle_call() as there is no cpuidle driver available to them by default. That causes a problem in fully loaded scenarios where CPUs go briefly idle for a couple of microseconds:

tick_nohz_idle_stop_tick() is invoked unconditionally which means unless there is timer pending in the next tick, the tick is stopped and a couple of microseconds later when the idle condition goes away restarted. That requires to program the clockevent device twice which implies a VM exit for each reprogramming.

It was suggested to remove the tick_nohz_idle_stop_tick() invocation from the default idle code, but would be counterproductive. It would not allow the host to go into deeper idle states when the guest CPU is fully idle as it has to maintain the periodic tick.

Cure this by implementing a trivial moving average filter which keeps track of the recent idle recidency time and only stop the tick when the average is larger than a tick.

Quoted 3 lines

Quoted 1 line

Signed-off-by Thomas Gleixner <[email protected]>

Shouldn't there be instead a new dedicated cpuidle driver with proper governor support?

I think a dummy cpuidle driver is an option, but calling into any governor seems overkill IMO, it presents an option to the user where there really is none (after all the cpuidle governor would just make a boolean decision as there are no states).

Quoted 25 lines · Christian Loehle

Quoted 23 lines · Frederic Weisbecker

Quoted 15 lines · Thomas Gleixner

Guests fall back to default_idle_call() as there is no cpuidle driver available to them by default. That causes a problem in fully loaded scenarios where CPUs go briefly idle for a couple of microseconds:

tick_nohz_idle_stop_tick() is invoked unconditionally which means unless there is timer pending in the next tick, the tick is stopped and a couple of microseconds later when the idle condition goes away restarted. That requires to program the clockevent device twice which implies a VM exit for each reprogramming.

It was suggested to remove the tick_nohz_idle_stop_tick() invocation from the default idle code, but would be counterproductive. It would not allow the host to go into deeper idle states when the guest CPU is fully idle as it has to maintain the periodic tick.

Quoted 5 lines

Cure this by implementing a trivial moving average filter which keeps track of the recent idle recidency time and only stop the tick when the average is larger than a tick.

Signed-off-by Thomas Gleixner <[email protected]>

Shouldn't there be instead a new dedicated cpuidle driver with proper governor support?

Quoted 4 lines

I think a dummy cpuidle driver is an option, but calling into any governor seems overkill IMO, it presents an option to the user where there really is none (after all the cpuidle governor would just make a boolean decision as there are no states).

I must confess I don't fully understand the picture with the non-existent states but what Thomas is doing in his patch is basically an ad-hoc implementation of cpuidle governor decision whether or not to stop the tick.

Quoted 26 lines · Frederic Weisbecker

Quoted 25 lines · Christian Loehle

Quoted 23 lines · Frederic Weisbecker

Quoted 15 lines · Thomas Gleixner

Guests fall back to default_idle_call() as there is no cpuidle driver available to them by default. That causes a problem in fully loaded scenarios where CPUs go briefly idle for a couple of microseconds:

tick_nohz_idle_stop_tick() is invoked unconditionally which means unless there is timer pending in the next tick, the tick is stopped and a couple of microseconds later when the idle condition goes away restarted. That requires to program the clockevent device twice which implies a VM exit for each reprogramming.

It was suggested to remove the tick_nohz_idle_stop_tick() invocation from the default idle code, but would be counterproductive. It would not allow the host to go into deeper idle states when the guest CPU is fully idle as it has to maintain the periodic tick.

Quoted 5 lines

Cure this by implementing a trivial moving average filter which keeps track of the recent idle recidency time and only stop the tick when the average is larger than a tick.

Signed-off-by Thomas Gleixner <[email protected]>

Shouldn't there be instead a new dedicated cpuidle driver with proper governor support?

Quoted 9 lines

Quoted 4 lines

I think a dummy cpuidle driver is an option, but calling into any governor seems overkill IMO, it presents an option to the user where there really is none (after all the cpuidle governor would just make a boolean decision as there are no states).

I must confess I don't fully understand the picture with the non-existent states but what Thomas is doing in his patch is basically an ad-hoc implementation of cpuidle governor decision whether or not to stop the tick.

Yup and if we put that into the cpuidle governor then we have to duplicate that logic for all governors even though for <= 1 states they hopefully should be the same.

A dummy driver would allow for this logic to live in drivers/cpuidle/ but I don't have a preference either way.

Guests fall back to default_idle_call() as there is no cpuidle driver available to them by default. That causes a problem in fully loaded scenarios where CPUs go briefly idle for a couple of microseconds:

tick_nohz_idle_stop_tick() is invoked unconditionally which means unless there is timer pending in the next tick, the tick is stopped and a couple of microseconds later when the idle condition goes away restarted. That requires to program the clockevent device twice which implies a VM exit for each reprogramming.

It was suggested to remove the tick_nohz_idle_stop_tick() invocation from the default idle code, but would be counterproductive. It would not allow the host to go into deeper idle states when the guest CPU is fully idle as it has to maintain the periodic tick.

Cure this by implementing a trivial moving average filter which keeps track of the recent idle recidency time and only stop the tick when the average is larger than a tick.

Signed-off-by Thomas Gleixner <[email protected]>

---
 kernel/sched/idle.c |   65 +++++++++++++++++++++++++++++++++++++++++++++-------
 1 file changed, 57 insertions(+), 8 deletions(-)

kernel/sched/idle.c +56 -8

@@ -105,12 +105,7 @@ static inline void cond_tick_broadcast_e
 static inline void cond_tick_broadcast_exit(void) { }
 #endif /* !CONFIG_GENERIC_CLOCKEVENTS_BROADCAST_IDLE */
 
-/**
- * default_idle_call - Default CPU idle routine.
- *
- * To use when the cpuidle framework cannot be used.
- */
-static void __cpuidle default_idle_call(void)
+static void __cpuidle __default_idle_call(void)
 {
 	instrumentation_begin();
 	if (!current_clr_polling_and_test()) {

@@ -130,6 +125,61 @@ static void __cpuidle default_idle_call(
 	instrumentation_end();
 }
 
+#ifdef CONFIG_NO_HZ_COMMON
+
+/* Limit to 4 entries so it fits in a cache line */
+#define IDLE_DUR_ENTRIES	4
+#define IDLE_DUR_MASK		(IDLE_DUR_ENTRIES - 1)
+
+struct idle_nohz_data {
+	u64		duration[IDLE_DUR_ENTRIES];
+	u64		entry_time;
+	u64		sum;
+	unsigned int	idx;
+};
+
+static DEFINE_PER_CPU_ALIGNED(struct idle_nohz_data, nohz_data);
+
+/**
+ * default_idle_call - Default CPU idle routine.
+ *
+ * To use when the cpuidle framework cannot be used.
+ */
+static void default_idle_call(void)
+{
+	struct idle_nohz_data *nd = this_cpu_ptr(&nohz_data);
+	unsigned int idx = nd->idx;
+	s64 delta;
+
+	/*
+	 * If the CPU spends more than a tick on average in idle, try to stop
+	 * the tick.
+	 */
+	if (nd->sum > TICK_NSEC * IDLE_DUR_ENTRIES)
+		tick_nohz_idle_stop_tick();
+
+	__default_idle_call();
+
+	/*
+	 * Build a moving average of the time spent in idle to prevent stopping
+	 * the tick on a loaded system which only goes idle briefly.
+	 */
+	delta = max(sched_clock() - nd->entry_time, 0);
+	nd->sum += delta - nd->duration[idx];
+	nd->duration[idx] = delta;
+	nd->idx = (idx + 1) & IDLE_DUR_MASK;
+}
+
+static void default_idle_enter(void)
+{
+	this_cpu_write(nohz_data.entry_time, sched_clock());
+}
+
+#else  /* CONFIG_NO_HZ_COMMON */
+static inline void default_idle_call(void { __default_idle_call(); }
+static inline void default_idle_enter(void) { }
+#endif /* !CONFIG_NO_HZ_COMMON */
+
 static int call_cpuidle_s2idle(struct cpuidle_driver *drv,
 			       struct cpuidle_device *dev,
 			       u64 max_latency_ns)

@@ -186,8 +236,6 @@ static void cpuidle_idle_call(void)
 	}
 
 	if (cpuidle_not_available(drv, dev)) {
-		tick_nohz_idle_stop_tick();
-
 		default_idle_call();
 		goto exit_idle;
 	}

Quoted 117 lines · Thomas Gleixner

Guests fall back to default_idle_call() as there is no cpuidle driver available to them by default. That causes a problem in fully loaded scenarios where CPUs go briefly idle for a couple of microseconds:

tick_nohz_idle_stop_tick() is invoked unconditionally which means unless there is timer pending in the next tick, the tick is stopped and a couple of microseconds later when the idle condition goes away restarted. That requires to program the clockevent device twice which implies a VM exit for each reprogramming.

It was suggested to remove the tick_nohz_idle_stop_tick() invocation from the default idle code, but would be counterproductive. It would not allow the host to go into deeper idle states when the guest CPU is fully idle as it has to maintain the periodic tick.

Cure this by implementing a trivial moving average filter which keeps track of the recent idle recidency time and only stop the tick when the average is larger than a tick.

Signed-off-by Thomas Gleixner <[email protected]>

---
 kernel/sched/idle.c |   65 +++++++++++++++++++++++++++++++++++++++++++++-------
 1 file changed, 57 insertions(+), 8 deletions(-)

kernel/sched/idle.c +57 -8

@@ -105,12 +105,7 @@ static inline void cond_tick_broadcast_e
 static inline void cond_tick_broadcast_exit(void) { }
 #endif /* !CONFIG_GENERIC_CLOCKEVENTS_BROADCAST_IDLE */
 
-/**
- * default_idle_call - Default CPU idle routine.
- *
- * To use when the cpuidle framework cannot be used.
- */
-static void __cpuidle default_idle_call(void)
+static void __cpuidle __default_idle_call(void)
 {
      instrumentation_begin();
      if (!current_clr_polling_and_test()) {

@@ -130,6 +125,61 @@ static void __cpuidle default_idle_call(
      instrumentation_end();
 }
 
+#ifdef CONFIG_NO_HZ_COMMON
+
+/* Limit to 4 entries so it fits in a cache line */
+#define IDLE_DUR_ENTRIES     4
+#define IDLE_DUR_MASK                (IDLE_DUR_ENTRIES - 1)
+
+struct idle_nohz_data {
+     u64             duration[IDLE_DUR_ENTRIES];
+     u64             entry_time;
+     u64             sum;
+     unsigned int    idx;
+};
+
+static DEFINE_PER_CPU_ALIGNED(struct idle_nohz_data, nohz_data);
+
+/**
+ * default_idle_call - Default CPU idle routine.
+ *
+ * To use when the cpuidle framework cannot be used.
+ */
+static void default_idle_call(void)
+{
+     struct idle_nohz_data *nd = this_cpu_ptr(&nohz_data);
+     unsigned int idx = nd->idx;
+     s64 delta;
+
+     /*
+      * If the CPU spends more than a tick on average in idle, try to stop
+      * the tick.
+      */
+     if (nd->sum > TICK_NSEC * IDLE_DUR_ENTRIES)
+             tick_nohz_idle_stop_tick();
+
+     __default_idle_call();
+
+     /*
+      * Build a moving average of the time spent in idle to prevent stopping
+      * the tick on a loaded system which only goes idle briefly.
+      */
+     delta = max(sched_clock() - nd->entry_time, 0);
+     nd->sum += delta - nd->duration[idx];
+     nd->duration[idx] = delta;
+     nd->idx = (idx + 1) & IDLE_DUR_MASK;
+}
+
+static void default_idle_enter(void)
+{
+     this_cpu_write(nohz_data.entry_time, sched_clock());
+}
+
+#else  /* CONFIG_NO_HZ_COMMON */
+static inline void default_idle_call(void { __default_idle_call(); }
+static inline void default_idle_enter(void) { }
+#endif /* !CONFIG_NO_HZ_COMMON */
+
 static int call_cpuidle_s2idle(struct cpuidle_driver *drv,
                             struct cpuidle_device *dev,
                             u64 max_latency_ns)

@@ -186,8 +236,6 @@ static void cpuidle_idle_call(void)
      }
 
      if (cpuidle_not_available(drv, dev)) {
-             tick_nohz_idle_stop_tick();
-
              default_idle_call();
              goto exit_idle;
      }

@@ -276,6 +324,7 @@ static void do_idle(void)
 
      __current_set_polling();
      tick_nohz_idle_enter();
+     default_idle_enter();
 

kernel/sched/idle.c static void do_idle(void) · line 330

Quoted 3 lines

kernel/sched/idle.c +0 -0

      while (!need_resched()) {
 

How does this work? We don't stop the tick until the average idle time is larger, but if we don't stop the tick how is that possible?

Why don't we just require one or two consecutive tick wakeups before stopping?

Quoted 126 lines · Christian Loehle

Quoted 113 lines · Thomas Gleixner

Guests fall back to default_idle_call() as there is no cpuidle driver available to them by default. That causes a problem in fully loaded scenarios where CPUs go briefly idle for a couple of microseconds:

tick_nohz_idle_stop_tick() is invoked unconditionally which means unless there is timer pending in the next tick, the tick is stopped and a couple of microseconds later when the idle condition goes away restarted. That requires to program the clockevent device twice which implies a VM exit for each reprogramming.

It was suggested to remove the tick_nohz_idle_stop_tick() invocation from the default idle code, but would be counterproductive. It would not allow the host to go into deeper idle states when the guest CPU is fully idle as it has to maintain the periodic tick.

Cure this by implementing a trivial moving average filter which keeps track of the recent idle recidency time and only stop the tick when the average is larger than a tick.

Signed-off-by Thomas Gleixner <[email protected]>

---
 kernel/sched/idle.c |   65 +++++++++++++++++++++++++++++++++++++++++++++-------
 1 file changed, 57 insertions(+), 8 deletions(-)

kernel/sched/idle.c +56 -8

@@ -105,12 +105,7 @@ static inline void cond_tick_broadcast_e
 static inline void cond_tick_broadcast_exit(void) { }
 #endif /* !CONFIG_GENERIC_CLOCKEVENTS_BROADCAST_IDLE */
 
-/**
- * default_idle_call - Default CPU idle routine.
- *
- * To use when the cpuidle framework cannot be used.
- */
-static void __cpuidle default_idle_call(void)
+static void __cpuidle __default_idle_call(void)
 {
      instrumentation_begin();
      if (!current_clr_polling_and_test()) {

@@ -130,6 +125,61 @@ static void __cpuidle default_idle_call(
      instrumentation_end();
 }
 
+#ifdef CONFIG_NO_HZ_COMMON
+
+/* Limit to 4 entries so it fits in a cache line */
+#define IDLE_DUR_ENTRIES     4
+#define IDLE_DUR_MASK                (IDLE_DUR_ENTRIES - 1)
+
+struct idle_nohz_data {
+     u64             duration[IDLE_DUR_ENTRIES];
+     u64             entry_time;
+     u64             sum;
+     unsigned int    idx;
+};
+
+static DEFINE_PER_CPU_ALIGNED(struct idle_nohz_data, nohz_data);
+
+/**
+ * default_idle_call - Default CPU idle routine.
+ *
+ * To use when the cpuidle framework cannot be used.
+ */
+static void default_idle_call(void)
+{
+     struct idle_nohz_data *nd = this_cpu_ptr(&nohz_data);
+     unsigned int idx = nd->idx;
+     s64 delta;
+
+     /*
+      * If the CPU spends more than a tick on average in idle, try to stop
+      * the tick.
+      */
+     if (nd->sum > TICK_NSEC * IDLE_DUR_ENTRIES)
+             tick_nohz_idle_stop_tick();
+
+     __default_idle_call();
+
+     /*
+      * Build a moving average of the time spent in idle to prevent stopping
+      * the tick on a loaded system which only goes idle briefly.
+      */
+     delta = max(sched_clock() - nd->entry_time, 0);
+     nd->sum += delta - nd->duration[idx];
+     nd->duration[idx] = delta;
+     nd->idx = (idx + 1) & IDLE_DUR_MASK;
+}
+
+static void default_idle_enter(void)
+{
+     this_cpu_write(nohz_data.entry_time, sched_clock());
+}
+
+#else  /* CONFIG_NO_HZ_COMMON */
+static inline void default_idle_call(void { __default_idle_call(); }
+static inline void default_idle_enter(void) { }
+#endif /* !CONFIG_NO_HZ_COMMON */
+
 static int call_cpuidle_s2idle(struct cpuidle_driver *drv,
                             struct cpuidle_device *dev,
                             u64 max_latency_ns)

@@ -186,8 +236,6 @@ static void cpuidle_idle_call(void)
      }
 
      if (cpuidle_not_available(drv, dev)) {
-             tick_nohz_idle_stop_tick();
-
              default_idle_call();
              goto exit_idle;
      }

kernel/sched/idle.c static void do_idle(void) · line 330

Quoted 7 lines

kernel/sched/idle.c +1 -0

@@ -276,6 +324,7 @@ static void do_idle(void)
 
      __current_set_polling();
      tick_nohz_idle_enter();
+     default_idle_enter();
 
      while (!need_resched()) {
 

How does this work? We don't stop the tick until the average idle time is larger, but if we don't stop the tick how is that possible?

Quoted 1 line

Why don't we just require one or two consecutive tick wakeups before stopping?

Exactly my thought and I think one should be sufficient.

Quoted 128 lines · Rafael J. Wysocki

Quoted 126 lines · Christian Loehle

Quoted 113 lines · Thomas Gleixner

Guests fall back to default_idle_call() as there is no cpuidle driver available to them by default. That causes a problem in fully loaded scenarios where CPUs go briefly idle for a couple of microseconds:

tick_nohz_idle_stop_tick() is invoked unconditionally which means unless there is timer pending in the next tick, the tick is stopped and a couple of microseconds later when the idle condition goes away restarted. That requires to program the clockevent device twice which implies a VM exit for each reprogramming.

It was suggested to remove the tick_nohz_idle_stop_tick() invocation from the default idle code, but would be counterproductive. It would not allow the host to go into deeper idle states when the guest CPU is fully idle as it has to maintain the periodic tick.

Cure this by implementing a trivial moving average filter which keeps track of the recent idle recidency time and only stop the tick when the average is larger than a tick.

Signed-off-by Thomas Gleixner <[email protected]>

---
 kernel/sched/idle.c |   65 +++++++++++++++++++++++++++++++++++++++++++++-------
 1 file changed, 57 insertions(+), 8 deletions(-)

kernel/sched/idle.c +56 -8

@@ -105,12 +105,7 @@ static inline void cond_tick_broadcast_e
 static inline void cond_tick_broadcast_exit(void) { }
 #endif /* !CONFIG_GENERIC_CLOCKEVENTS_BROADCAST_IDLE */
 
-/**
- * default_idle_call - Default CPU idle routine.
- *
- * To use when the cpuidle framework cannot be used.
- */
-static void __cpuidle default_idle_call(void)
+static void __cpuidle __default_idle_call(void)
 {
      instrumentation_begin();
      if (!current_clr_polling_and_test()) {

@@ -130,6 +125,61 @@ static void __cpuidle default_idle_call(
      instrumentation_end();
 }
 
+#ifdef CONFIG_NO_HZ_COMMON
+
+/* Limit to 4 entries so it fits in a cache line */
+#define IDLE_DUR_ENTRIES     4
+#define IDLE_DUR_MASK                (IDLE_DUR_ENTRIES - 1)
+
+struct idle_nohz_data {
+     u64             duration[IDLE_DUR_ENTRIES];
+     u64             entry_time;
+     u64             sum;
+     unsigned int    idx;
+};
+
+static DEFINE_PER_CPU_ALIGNED(struct idle_nohz_data, nohz_data);
+
+/**
+ * default_idle_call - Default CPU idle routine.
+ *
+ * To use when the cpuidle framework cannot be used.
+ */
+static void default_idle_call(void)
+{
+     struct idle_nohz_data *nd = this_cpu_ptr(&nohz_data);
+     unsigned int idx = nd->idx;
+     s64 delta;
+
+     /*
+      * If the CPU spends more than a tick on average in idle, try to stop
+      * the tick.
+      */
+     if (nd->sum > TICK_NSEC * IDLE_DUR_ENTRIES)
+             tick_nohz_idle_stop_tick();
+
+     __default_idle_call();
+
+     /*
+      * Build a moving average of the time spent in idle to prevent stopping
+      * the tick on a loaded system which only goes idle briefly.
+      */
+     delta = max(sched_clock() - nd->entry_time, 0);
+     nd->sum += delta - nd->duration[idx];
+     nd->duration[idx] = delta;
+     nd->idx = (idx + 1) & IDLE_DUR_MASK;
+}
+
+static void default_idle_enter(void)
+{
+     this_cpu_write(nohz_data.entry_time, sched_clock());
+}
+
+#else  /* CONFIG_NO_HZ_COMMON */
+static inline void default_idle_call(void { __default_idle_call(); }
+static inline void default_idle_enter(void) { }
+#endif /* !CONFIG_NO_HZ_COMMON */
+
 static int call_cpuidle_s2idle(struct cpuidle_driver *drv,
                             struct cpuidle_device *dev,
                             u64 max_latency_ns)

@@ -186,8 +236,6 @@ static void cpuidle_idle_call(void)
      }
 
      if (cpuidle_not_available(drv, dev)) {
-             tick_nohz_idle_stop_tick();
-
              default_idle_call();
              goto exit_idle;
      }

kernel/sched/idle.c static void do_idle(void) · line 330

Quoted 7 lines

kernel/sched/idle.c +1 -0

@@ -276,6 +324,7 @@ static void do_idle(void)
 
      __current_set_polling();
      tick_nohz_idle_enter();
+     default_idle_enter();
 
      while (!need_resched()) {
 

How does this work? We don't stop the tick until the average idle time is larger, but if we don't stop the tick how is that possible?

Quoted 3 lines

Quoted 1 line

Why don't we just require one or two consecutive tick wakeups before stopping?

Exactly my thought and I think one should be sufficient.

I concur. From our experience with TEO util threshold these averages can backfire. I think one tick is sufficient delay to not be obviously broken.

Quoted 129 lines · CET Qais Yousef

Quoted 128 lines · Rafael J. Wysocki

Quoted 126 lines · Christian Loehle

Quoted 113 lines · Thomas Gleixner

Guests fall back to default_idle_call() as there is no cpuidle driver available to them by default. That causes a problem in fully loaded scenarios where CPUs go briefly idle for a couple of microseconds:

tick_nohz_idle_stop_tick() is invoked unconditionally which means unless there is timer pending in the next tick, the tick is stopped and a couple of microseconds later when the idle condition goes away restarted. That requires to program the clockevent device twice which implies a VM exit for each reprogramming.

It was suggested to remove the tick_nohz_idle_stop_tick() invocation from the default idle code, but would be counterproductive. It would not allow the host to go into deeper idle states when the guest CPU is fully idle as it has to maintain the periodic tick.

Cure this by implementing a trivial moving average filter which keeps track of the recent idle recidency time and only stop the tick when the average is larger than a tick.

Signed-off-by Thomas Gleixner <[email protected]>

---
 kernel/sched/idle.c |   65 +++++++++++++++++++++++++++++++++++++++++++++-------
 1 file changed, 57 insertions(+), 8 deletions(-)

kernel/sched/idle.c +56 -8

@@ -105,12 +105,7 @@ static inline void cond_tick_broadcast_e
 static inline void cond_tick_broadcast_exit(void) { }
 #endif /* !CONFIG_GENERIC_CLOCKEVENTS_BROADCAST_IDLE */
 
-/**
- * default_idle_call - Default CPU idle routine.
- *
- * To use when the cpuidle framework cannot be used.
- */
-static void __cpuidle default_idle_call(void)
+static void __cpuidle __default_idle_call(void)
 {
      instrumentation_begin();
      if (!current_clr_polling_and_test()) {

@@ -130,6 +125,61 @@ static void __cpuidle default_idle_call(
      instrumentation_end();
 }
 
+#ifdef CONFIG_NO_HZ_COMMON
+
+/* Limit to 4 entries so it fits in a cache line */
+#define IDLE_DUR_ENTRIES     4
+#define IDLE_DUR_MASK                (IDLE_DUR_ENTRIES - 1)
+
+struct idle_nohz_data {
+     u64             duration[IDLE_DUR_ENTRIES];
+     u64             entry_time;
+     u64             sum;
+     unsigned int    idx;
+};
+
+static DEFINE_PER_CPU_ALIGNED(struct idle_nohz_data, nohz_data);
+
+/**
+ * default_idle_call - Default CPU idle routine.
+ *
+ * To use when the cpuidle framework cannot be used.
+ */
+static void default_idle_call(void)
+{
+     struct idle_nohz_data *nd = this_cpu_ptr(&nohz_data);
+     unsigned int idx = nd->idx;
+     s64 delta;
+
+     /*
+      * If the CPU spends more than a tick on average in idle, try to stop
+      * the tick.
+      */
+     if (nd->sum > TICK_NSEC * IDLE_DUR_ENTRIES)
+             tick_nohz_idle_stop_tick();
+
+     __default_idle_call();
+
+     /*
+      * Build a moving average of the time spent in idle to prevent stopping
+      * the tick on a loaded system which only goes idle briefly.
+      */
+     delta = max(sched_clock() - nd->entry_time, 0);
+     nd->sum += delta - nd->duration[idx];
+     nd->duration[idx] = delta;
+     nd->idx = (idx + 1) & IDLE_DUR_MASK;
+}
+
+static void default_idle_enter(void)
+{
+     this_cpu_write(nohz_data.entry_time, sched_clock());
+}
+
+#else  /* CONFIG_NO_HZ_COMMON */
+static inline void default_idle_call(void { __default_idle_call(); }
+static inline void default_idle_enter(void) { }
+#endif /* !CONFIG_NO_HZ_COMMON */
+
 static int call_cpuidle_s2idle(struct cpuidle_driver *drv,
                             struct cpuidle_device *dev,
                             u64 max_latency_ns)

@@ -186,8 +236,6 @@ static void cpuidle_idle_call(void)
      }
 
      if (cpuidle_not_available(drv, dev)) {
-             tick_nohz_idle_stop_tick();
-
              default_idle_call();
              goto exit_idle;
      }

kernel/sched/idle.c static void do_idle(void) · line 330

Quoted 7 lines

kernel/sched/idle.c +1 -0

@@ -276,6 +324,7 @@ static void do_idle(void)
 
      __current_set_polling();
      tick_nohz_idle_enter();
+     default_idle_enter();
 
      while (!need_resched()) {
 

How does this work? We don't stop the tick until the average idle time is larger, but if we don't stop the tick how is that possible?

Quoted 6 lines

Quoted 3 lines

Quoted 1 line

Why don't we just require one or two consecutive tick wakeups before stopping?

Exactly my thought and I think one should be sufficient.

I concur. From our experience with TEO util threshold these averages can backfire. I think one tick is sufficient delay to not be obviously broken.

So if I'm not mistaken, it would be something like the appended prototype (completely untested, but it builds for me).

---
 drivers/cpuidle/cpuidle.c |   10 ----------
 kernel/sched/idle.c       |   32 ++++++++++++++++++++++++--------
 2 files changed, 24 insertions(+), 18 deletions(-)

drivers/cpuidle/cpuidle.c +0 -10

@@ -359,16 +359,6 @@ noinstr int cpuidle_enter_state(struct c
 int cpuidle_select(struct cpuidle_driver *drv, struct cpuidle_device *dev,
                   bool *stop_tick)
 {
-       /*
-        * If there is only a single idle state (or none), there is nothing
-        * meaningful for the governor to choose. Skip the governor and
-        * always use state 0 with the tick running.
-        */
-       if (drv->state_count <= 1) {
-               *stop_tick = false;
-               return 0;
-       }
-
        return cpuidle_curr_governor->select(drv, dev, stop_tick);
 }
 

kernel/sched/idle.c +9 -1

@@ -161,6 +161,14 @@ static int call_cpuidle(struct cpuidle_d
        return cpuidle_enter(drv, dev, next_state);
 }
 
+static void idle_call_stop_or_retain_tick(bool stop_tick)
+{
+       if (stop_tick || tick_nohz_tick_stopped())
+               tick_nohz_idle_stop_tick();
+       else
+               tick_nohz_idle_retain_tick();
+}
+
 /**
  * cpuidle_idle_call - the main idle function
  *

@@ -170,7 +178,7 @@ static int call_cpuidle(struct cpuidle_d
  * set, and it returns with polling set.  If it ever stops polling, it
  * must clear the polling bit.
  */
-static void cpuidle_idle_call(void)
+static void cpuidle_idle_call(bool got_tick)
 {
        struct cpuidle_device *dev = cpuidle_get_device();
        struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);

kernel/sched/idle.c +14 -6

                default_idle_call();
                goto exit_idle;

@@ -221,7 +229,7 @@ static void cpuidle_idle_call(void)
 
                next_state = cpuidle_find_deepest_state(drv, dev, max_latency_ns);
                call_cpuidle(drv, dev, next_state);
-       } else {
+       } else if (drv->state_count > 1) {
                bool stop_tick = true;
 
                /*

@@ -229,16 +237,22 @@ static void cpuidle_idle_call(void)
                 */
                next_state = cpuidle_select(drv, dev, &stop_tick);
 
-               if (stop_tick || tick_nohz_tick_stopped())
-                       tick_nohz_idle_stop_tick();
-               else
-                       tick_nohz_idle_retain_tick();
+               idle_call_stop_or_retain_tick(stop_tick);
 
                entered_state = call_cpuidle(drv, dev, next_state);
                /*
                 * Give the governor an opportunity to reflect on the outcome
                 */
                cpuidle_reflect(dev, entered_state);
+       } else {
+               /*
+                * If there is only a single idle state (or none), there is
+                * nothing meaningful for the governor to choose.  Skip the
+                * governor and always use state 0.
+                */
+               idle_call_stop_or_retain_tick(!got_tick);
+
+               call_cpuidle(drv, dev, 0);
        }
 
 exit_idle:

@@ -259,6 +273,7 @@ exit_idle:
 static void do_idle(void)
 {
        int cpu = smp_processor_id();
+       bool got_tick = false;
 
        /*
         * Check if we need to update blocked load

@@ -329,8 +344,9 @@ static void do_idle(void)
                        tick_nohz_idle_restart_tick();
                        cpu_idle_poll();
                } else {
-                       cpuidle_idle_call();
+                       cpuidle_idle_call(got_tick);
                }
+               got_tick = tick_nohz_idle_got_tick();
                arch_cpu_idle_exit();
        }

Quoted 3 lines · Rafael J. Wysocki

Quoted 2 lines · CET Qais Yousef

Quoted 1 line · Rafael J. Wysocki

On Mon, Mar 2, 2026 at 12:04 PM Christian Loehle

Quoted 64 lines

Quoted 7 lines

Quoted 4 lines

Quoted 2 lines

Why don't we just require one or two consecutive tick wakeups before stopping?

Exactly my thought and I think one should be sufficient.

I concur. From our experience with TEO util threshold these averages can backfire. I think one tick is sufficient delay to not be obviously broken.

So if I'm not mistaken, it would be something like the appended prototype (completely untested, but it builds for me).

---
 drivers/cpuidle/cpuidle.c |   10 ----------
 kernel/sched/idle.c       |   32 ++++++++++++++++++++++++--------
 2 files changed, 24 insertions(+), 18 deletions(-)

drivers/cpuidle/cpuidle.c +0 -10

@@ -359,16 +359,6 @@ noinstr int cpuidle_enter_state(struct c
 int cpuidle_select(struct cpuidle_driver *drv, struct cpuidle_device *dev,
                   bool *stop_tick)
 {
-       /*
-        * If there is only a single idle state (or none), there is nothing
-        * meaningful for the governor to choose. Skip the governor and
-        * always use state 0 with the tick running.
-        */
-       if (drv->state_count <= 1) {
-               *stop_tick = false;
-               return 0;
-       }
-
        return cpuidle_curr_governor->select(drv, dev, stop_tick);
 }
 

kernel/sched/idle.c +9 -1

@@ -161,6 +161,14 @@ static int call_cpuidle(struct cpuidle_d
        return cpuidle_enter(drv, dev, next_state);
 }
 
+static void idle_call_stop_or_retain_tick(bool stop_tick)
+{
+       if (stop_tick || tick_nohz_tick_stopped())
+               tick_nohz_idle_stop_tick();
+       else
+               tick_nohz_idle_retain_tick();
+}
+
 /**
  * cpuidle_idle_call - the main idle function
  *

@@ -170,7 +178,7 @@ static int call_cpuidle(struct cpuidle_d
  * set, and it returns with polling set.  If it ever stops polling, it
  * must clear the polling bit.
  */
-static void cpuidle_idle_call(void)
+static void cpuidle_idle_call(bool got_tick)
 {
        struct cpuidle_device *dev = cpuidle_get_device();
        struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);

@@ -186,7 +194,7 @@ static void cpuidle_idle_call(void)
        }
 

kernel/sched/idle.c static void cpuidle_idle_call(void) · line 197

Quoted 3 lines

kernel/sched/idle.c +1 -1

        if (cpuidle_not_available(drv, dev)) {
-               tick_nohz_idle_stop_tick();
+               idle_call_stop_or_retain_tick(!got_tick);

Oh, I got this backwards (here and below).

The tick should be stopped if we've got the tick previously, but you get the idea.

In the meantime I realized that if the .select() governor callback is skipped, its .reflect() callback should be skipped either, so I've posted this:

https://lkml.org/lkml/2026/3/7/569

and here's a fixed version of the last patch on top of the above (for completeness):

---
 kernel/sched/idle.c |   25 ++++++++++++++++---------
 1 file changed, 16 insertions(+), 9 deletions(-)

kernel/sched/idle.c +10 -2

@@ -161,6 +161,14 @@ static int call_cpuidle(struct cpuidle_d
 	return cpuidle_enter(drv, dev, next_state);
 }
 
+static void idle_call_stop_or_retain_tick(bool stop_tick)
+{
+	if (stop_tick || tick_nohz_tick_stopped())
+		tick_nohz_idle_stop_tick();
+	else
+		tick_nohz_idle_retain_tick();
+}
+
 /**
  * cpuidle_idle_call - the main idle function
  *

@@ -170,7 +178,7 @@ static int call_cpuidle(struct cpuidle_d
  * set, and it returns with polling set.  If it ever stops polling, it
  * must clear the polling bit.
  */
-static void cpuidle_idle_call(void)
+static void cpuidle_idle_call(bool stop_tick)
 {
 	struct cpuidle_device *dev = cpuidle_get_device();
 	struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);

@@ -186,7 +194,7 @@ static void cpuidle_idle_call(void)
 	}
 
 	if (cpuidle_not_available(drv, dev)) {
-		tick_nohz_idle_stop_tick();
+		idle_call_stop_or_retain_tick(stop_tick);
 
 		default_idle_call();
 		goto exit_idle;

kernel/sched/idle.c +5 -6

 		/*
 		 * Ask the cpuidle framework to choose a convenient idle state.
 		 */
 		next_state = cpuidle_select(drv, dev, &stop_tick);
 
-		if (stop_tick || tick_nohz_tick_stopped())
-			tick_nohz_idle_stop_tick();
-		else
-			tick_nohz_idle_retain_tick();
+		idle_call_stop_or_retain_tick(stop_tick);
 
 		entered_state = call_cpuidle(drv, dev, next_state);
 		/*

@@ -240,7 +245,7 @@ static void cpuidle_idle_call(void)
 		 */
 		cpuidle_reflect(dev, entered_state);
 	} else {
-		tick_nohz_idle_retain_tick();
+		idle_call_stop_or_retain_tick(stop_tick);
 
 		/*
 		 * If there is only a single idle state (or none), there is

@@ -268,6 +273,7 @@ exit_idle:
 static void do_idle(void)
 {
 	int cpu = smp_processor_id();
+	bool got_tick = false;
 
 	/*
 	 * Check if we need to update blocked load

@@ -338,8 +344,9 @@ static void do_idle(void)
 			tick_nohz_idle_restart_tick();
 			cpu_idle_poll();
 		} else {
-			cpuidle_idle_call();
+			cpuidle_idle_call(got_tick);
 		}
+		got_tick = tick_nohz_idle_got_tick();
 		arch_cpu_idle_exit();
 	}
 

Quoted 46 lines · Rafael J. Wysocki

In the meantime I realized that if the .select() governor callback is skipped, its .reflect() callback should be skipped either, so I've posted this:

https://lkml.org/lkml/2026/3/7/569

and here's a fixed version of the last patch on top of the above (for completeness):

---
 kernel/sched/idle.c |   25 ++++++++++++++++---------
 1 file changed, 16 insertions(+), 9 deletions(-)

kernel/sched/idle.c +10 -2

@@ -161,6 +161,14 @@ static int call_cpuidle(struct cpuidle_d
 	return cpuidle_enter(drv, dev, next_state);
 }
 
+static void idle_call_stop_or_retain_tick(bool stop_tick)
+{
+	if (stop_tick || tick_nohz_tick_stopped())
+		tick_nohz_idle_stop_tick();
+	else
+		tick_nohz_idle_retain_tick();
+}
+
 /**
  * cpuidle_idle_call - the main idle function
  *

@@ -170,7 +178,7 @@ static int call_cpuidle(struct cpuidle_d
  * set, and it returns with polling set.  If it ever stops polling, it
  * must clear the polling bit.
  */
-static void cpuidle_idle_call(void)
+static void cpuidle_idle_call(bool stop_tick)
 {
 	struct cpuidle_device *dev = cpuidle_get_device();
 	struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);

@@ -186,7 +194,7 @@ static void cpuidle_idle_call(void)
 	}
 
 	if (cpuidle_not_available(drv, dev)) {
-		tick_nohz_idle_stop_tick();
+		idle_call_stop_or_retain_tick(stop_tick);
 

kernel/sched/idle.c static void cpuidle_idle_call(void) · line 233

Quoted 8 lines

kernel/sched/idle.c +1 -1

 		default_idle_call();
 		goto exit_idle;

@@ -222,17 +230,14 @@ static void cpuidle_idle_call(void)
 		next_state = cpuidle_find_deepest_state(drv, dev, max_latency_ns);
 		call_cpuidle(drv, dev, next_state);
 	} else if (drv->state_count > 1) {
-		bool stop_tick = true;
+		stop_tick = true;

Silly question, but wouldn't this benefit the normal path too to delay for one tick? This will only matter for the cases where the governor doesn't explicitly set stop_tick to either true or false - which I am not sure what they are :)

Quoted 48 lines · Qais Yousef

Quoted 46 lines · Rafael J. Wysocki

In the meantime I realized that if the .select() governor callback is skipped, its .reflect() callback should be skipped either, so I've posted this:

https://lkml.org/lkml/2026/3/7/569

and here's a fixed version of the last patch on top of the above (for completeness):

---
 kernel/sched/idle.c |   25 ++++++++++++++++---------
 1 file changed, 16 insertions(+), 9 deletions(-)

kernel/sched/idle.c +10 -2

@@ -161,6 +161,14 @@ static int call_cpuidle(struct cpuidle_d
 	return cpuidle_enter(drv, dev, next_state);
 }
 
+static void idle_call_stop_or_retain_tick(bool stop_tick)
+{
+	if (stop_tick || tick_nohz_tick_stopped())
+		tick_nohz_idle_stop_tick();
+	else
+		tick_nohz_idle_retain_tick();
+}
+
 /**
  * cpuidle_idle_call - the main idle function
  *

@@ -170,7 +178,7 @@ static int call_cpuidle(struct cpuidle_d
  * set, and it returns with polling set.  If it ever stops polling, it
  * must clear the polling bit.
  */
-static void cpuidle_idle_call(void)
+static void cpuidle_idle_call(bool stop_tick)
 {
 	struct cpuidle_device *dev = cpuidle_get_device();
 	struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);

@@ -186,7 +194,7 @@ static void cpuidle_idle_call(void)
 	}
 
 	if (cpuidle_not_available(drv, dev)) {
-		tick_nohz_idle_stop_tick();
+		idle_call_stop_or_retain_tick(stop_tick);
 

Quoted 13 lines

kernel/sched/idle.c static void cpuidle_idle_call(void) · line 233

Quoted 8 lines

kernel/sched/idle.c +1 -1

 		default_idle_call();
 		goto exit_idle;

@@ -222,17 +230,14 @@ static void cpuidle_idle_call(void)
 		next_state = cpuidle_find_deepest_state(drv, dev, max_latency_ns);
 		call_cpuidle(drv, dev, next_state);
 	} else if (drv->state_count > 1) {
-		bool stop_tick = true;
+		stop_tick = true;

Silly question, but wouldn't this benefit the normal path too to delay for one tick? This will only matter for the cases where the governor doesn't explicitly set stop_tick to either true or false - which I am not sure what they are :)

Right now the governors will always set stop_tick explicitly (and overriding that might confuse the governor-internal state).

Quoted 62 lines · Christian Loehle

Quoted 61 lines · Qais Yousef

Quoted 49 lines · Rafael J. Wysocki

In the meantime I realized that if the .select() governor callback is skipped, its .reflect() callback should be skipped either, so I've posted this:

https://lkml.org/lkml/2026/3/7/569

and here's a fixed version of the last patch on top of the above (for completeness):

---
 kernel/sched/idle.c |   25 ++++++++++++++++---------
 1 file changed, 16 insertions(+), 9 deletions(-)

kernel/sched/idle.c +10 -2

@@ -161,6 +161,14 @@ static int call_cpuidle(struct cpuidle_d
   return cpuidle_enter(drv, dev, next_state);
 }
 
+static void idle_call_stop_or_retain_tick(bool stop_tick)
+{
+  if (stop_tick || tick_nohz_tick_stopped())
+          tick_nohz_idle_stop_tick();
+  else
+          tick_nohz_idle_retain_tick();
+}
+
 /**
  * cpuidle_idle_call - the main idle function
  *

@@ -170,7 +178,7 @@ static int call_cpuidle(struct cpuidle_d
  * set, and it returns with polling set.  If it ever stops polling, it
  * must clear the polling bit.
  */
-static void cpuidle_idle_call(void)
+static void cpuidle_idle_call(bool stop_tick)
 {
   struct cpuidle_device *dev = cpuidle_get_device();
   struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);

@@ -186,7 +194,7 @@ static void cpuidle_idle_call(void)
   }
 
   if (cpuidle_not_available(drv, dev)) {
-          tick_nohz_idle_stop_tick();
+          idle_call_stop_or_retain_tick(stop_tick);
 
           default_idle_call();
           goto exit_idle;

kernel/sched/idle.c static void cpuidle_idle_call(void) · line 233

Quoted 5 lines

kernel/sched/idle.c +1 -1

@@ -222,17 +230,14 @@ static void cpuidle_idle_call(void)
           next_state = cpuidle_find_deepest_state(drv, dev, max_latency_ns);
           call_cpuidle(drv, dev, next_state);
   } else if (drv->state_count > 1) {
-          bool stop_tick = true;
+          stop_tick = true;

Silly question, but wouldn't this benefit the normal path too to delay for one tick? This will only matter for the cases where the governor doesn't explicitly set stop_tick to either true or false - which I am not sure what they are :)

Quoted 2 lines

Right now the governors will always set stop_tick explicitly (and overriding that might confuse the governor-internal state).

So we can drop this hunk then

Quoted 67 lines · Qais Yousef

Quoted 64 lines · Christian Loehle

Quoted 48 lines · Qais Yousef

Quoted 46 lines · Rafael J. Wysocki

In the meantime I realized that if the .select() governor callback is skipped, its .reflect() callback should be skipped either, so I've posted this:

https://lkml.org/lkml/2026/3/7/569

and here's a fixed version of the last patch on top of the above (for completeness):

---
 kernel/sched/idle.c |   25 ++++++++++++++++---------
 1 file changed, 16 insertions(+), 9 deletions(-)

kernel/sched/idle.c +10 -2

@@ -161,6 +161,14 @@ static int call_cpuidle(struct cpuidle_d
   return cpuidle_enter(drv, dev, next_state);
 }
 
+static void idle_call_stop_or_retain_tick(bool stop_tick)
+{
+  if (stop_tick || tick_nohz_tick_stopped())
+          tick_nohz_idle_stop_tick();
+  else
+          tick_nohz_idle_retain_tick();
+}
+
 /**
  * cpuidle_idle_call - the main idle function
  *

@@ -170,7 +178,7 @@ static int call_cpuidle(struct cpuidle_d
  * set, and it returns with polling set.  If it ever stops polling, it
  * must clear the polling bit.
  */
-static void cpuidle_idle_call(void)
+static void cpuidle_idle_call(bool stop_tick)
 {
   struct cpuidle_device *dev = cpuidle_get_device();
   struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);

@@ -186,7 +194,7 @@ static void cpuidle_idle_call(void)
   }
 
   if (cpuidle_not_available(drv, dev)) {
-          tick_nohz_idle_stop_tick();
+          idle_call_stop_or_retain_tick(stop_tick);
 

Quoted 13 lines

kernel/sched/idle.c static void cpuidle_idle_call(void) · line 233

Quoted 8 lines

kernel/sched/idle.c +1 -1

           default_idle_call();
           goto exit_idle;

@@ -222,17 +230,14 @@ static void cpuidle_idle_call(void)
           next_state = cpuidle_find_deepest_state(drv, dev, max_latency_ns);
           call_cpuidle(drv, dev, next_state);
   } else if (drv->state_count > 1) {
-          bool stop_tick = true;
+          stop_tick = true;

Silly question, but wouldn't this benefit the normal path too to delay for one tick? This will only matter for the cases where the governor doesn't explicitly set stop_tick to either true or false - which I am not sure what they are :)

Right now the governors will always set stop_tick explicitly (and overriding that might confuse the governor-internal state).

Quoted 1 line

So we can drop this hunk then

Not really.

The governors expect that stop_tick is true by default and they clear it if needed/desired. They may be confused if it is false to start with (theoretically, a governor may select an idle state with target residency beyond the tick period length then which won't make sense).

From: Rafael J. Wysocki <[email protected]>

If the cpuidle governor .select() callback is skipped because there is only one idle state in the cpuidle driver, the .reflect() callback should be skipped as well, at least for consistency (if not for correctness), so do it.

Fixes e5c9ffc6ae1b ("cpuidle: Skip governor when only one idle state is available")

Signed-off-by Rafael J. Wysocki <[email protected]>

---
 drivers/cpuidle/cpuidle.c |   10 ----------
 kernel/sched/idle.c       |   11 ++++++++++-
 2 files changed, 10 insertions(+), 11 deletions(-)

drivers/cpuidle/cpuidle.c +0 -10

@@ -359,16 +359,6 @@ noinstr int cpuidle_enter_state(struct c
 int cpuidle_select(struct cpuidle_driver *drv, struct cpuidle_device *dev,
 		   bool *stop_tick)
 {
-	/*
-	 * If there is only a single idle state (or none), there is nothing
-	 * meaningful for the governor to choose. Skip the governor and
-	 * always use state 0 with the tick running.
-	 */
-	if (drv->state_count <= 1) {
-		*stop_tick = false;
-		return 0;
-	}
-
 	return cpuidle_curr_governor->select(drv, dev, stop_tick);
 }
 

kernel/sched/idle.c +3 -1

@@ -221,7 +221,7 @@ static void cpuidle_idle_call(void)
 
 		next_state = cpuidle_find_deepest_state(drv, dev, max_latency_ns);
 		call_cpuidle(drv, dev, next_state);
-	} else {
+	} else if (drv->state_count > 1) {
 		bool stop_tick = true;
 
 		/*

@@ -239,6 +239,15 @@ static void cpuidle_idle_call(void)
 		 * Give the governor an opportunity to reflect on the outcome
 		 */
 		cpuidle_reflect(dev, entered_state);
+	} else {
+		tick_nohz_idle_retain_tick();

Quoted 59 lines · Rafael J. Wysocki

From: Rafael J. Wysocki <[email protected]>

If the cpuidle governor .select() callback is skipped because there is only one idle state in the cpuidle driver, the .reflect() callback should be skipped as well, at least for consistency (if not for correctness), so do it.

Fixes e5c9ffc6ae1b ("cpuidle: Skip governor when only one idle state is available")

Signed-off-by Rafael J. Wysocki <[email protected]>

---
 drivers/cpuidle/cpuidle.c |   10 ----------
 kernel/sched/idle.c       |   11 ++++++++++-
 2 files changed, 10 insertions(+), 11 deletions(-)

drivers/cpuidle/cpuidle.c +0 -10

@@ -359,16 +359,6 @@ noinstr int cpuidle_enter_state(struct c
 int cpuidle_select(struct cpuidle_driver *drv, struct cpuidle_device *dev,
                 bool *stop_tick)
 {
-     /*
-      * If there is only a single idle state (or none), there is nothing
-      * meaningful for the governor to choose. Skip the governor and
-      * always use state 0 with the tick running.
-      */
-     if (drv->state_count <= 1) {
-             *stop_tick = false;
-             return 0;
-     }
-
      return cpuidle_curr_governor->select(drv, dev, stop_tick);
 }
 

kernel/sched/idle.c +10 -1

@@ -221,7 +221,7 @@ static void cpuidle_idle_call(void)
 
              next_state = cpuidle_find_deepest_state(drv, dev, max_latency_ns);
              call_cpuidle(drv, dev, next_state);
-     } else {
+     } else if (drv->state_count > 1) {
              bool stop_tick = true;
 
              /*

@@ -239,6 +239,15 @@ static void cpuidle_idle_call(void)
               * Give the governor an opportunity to reflect on the outcome
               */
              cpuidle_reflect(dev, entered_state);
+     } else {
+             tick_nohz_idle_retain_tick();
+
+             /*
+              * If there is only a single idle state (or none), there is
+              * nothing meaningful for the governor to choose.  Skip the
+              * governor and always use state 0.
+              */
+             call_cpuidle(drv, dev, 0);
      }
 

kernel/sched/idle.c static void cpuidle_idle_call(void) · line 253

Quoted 4 lines

kernel/sched/idle.c +0 -0

 exit_idle:

Duh, good catch.

Reviewed-by Christian Loehle <[email protected]>

Quoted 61 lines · Christian Loehle

Quoted 59 lines · Rafael J. Wysocki

From: Rafael J. Wysocki <[email protected]>

If the cpuidle governor .select() callback is skipped because there is only one idle state in the cpuidle driver, the .reflect() callback should be skipped as well, at least for consistency (if not for correctness), so do it.

Fixes e5c9ffc6ae1b ("cpuidle: Skip governor when only one idle state is available")

Signed-off-by Rafael J. Wysocki <[email protected]>

---
 drivers/cpuidle/cpuidle.c |   10 ----------
 kernel/sched/idle.c       |   11 ++++++++++-
 2 files changed, 10 insertions(+), 11 deletions(-)

drivers/cpuidle/cpuidle.c +0 -10

@@ -359,16 +359,6 @@ noinstr int cpuidle_enter_state(struct c
 int cpuidle_select(struct cpuidle_driver *drv, struct cpuidle_device *dev,
                 bool *stop_tick)
 {
-     /*
-      * If there is only a single idle state (or none), there is nothing
-      * meaningful for the governor to choose. Skip the governor and
-      * always use state 0 with the tick running.
-      */
-     if (drv->state_count <= 1) {
-             *stop_tick = false;
-             return 0;
-     }
-
      return cpuidle_curr_governor->select(drv, dev, stop_tick);
 }
 

kernel/sched/idle.c +10 -1

@@ -221,7 +221,7 @@ static void cpuidle_idle_call(void)
 
              next_state = cpuidle_find_deepest_state(drv, dev, max_latency_ns);
              call_cpuidle(drv, dev, next_state);
-     } else {
+     } else if (drv->state_count > 1) {
              bool stop_tick = true;
 
              /*

@@ -239,6 +239,15 @@ static void cpuidle_idle_call(void)
               * Give the governor an opportunity to reflect on the outcome
               */
              cpuidle_reflect(dev, entered_state);
+     } else {
+             tick_nohz_idle_retain_tick();
+
+             /*
+              * If there is only a single idle state (or none), there is
+              * nothing meaningful for the governor to choose.  Skip the
+              * governor and always use state 0.
+              */
+             call_cpuidle(drv, dev, 0);
      }
 

Quoted 7 lines

kernel/sched/idle.c static void cpuidle_idle_call(void) · line 253

Quoted 4 lines

kernel/sched/idle.c +0 -0

 exit_idle:

Duh, good catch.

Reviewed-by Christian Loehle <[email protected]>

OK, so any objections or concerns from anyone?

I'm about to queue this up for the next -rc.

From: Rafael J. Wysocki <[email protected]>

If the cpuidle governor .select() callback is skipped because there is only one idle state in the cpuidle driver, the .reflect() callback should be skipped as well, at least for consistency (if not for correctness), so do it.

Fixes e5c9ffc6ae1b ("cpuidle: Skip governor when only one idle state

is available")

Signed-off-by Rafael J. Wysocki <[email protected]>

---
 drivers/cpuidle/cpuidle.c |   10 ----------
 kernel/sched/idle.c       |   11 ++++++++++-

 2 files changed, 10 insertions(+), 11 deletions(-)

drivers/cpuidle/cpuidle.c +0 -0

@@ -359,16 +359,6 @@ noinstr int cpuidle_enter_state(struct c

 int cpuidle_select(struct cpuidle_driver *drv, struct cpuidle_device *dev,
     bool *stop_tick)
 {

drivers/cpuidle/cpuidle.c +0 -2

-	/*
-	 * If there is only a single idle state (or none), there is

nothing

drivers/cpuidle/cpuidle.c +0 -1

-	 * meaningful for the governor to choose. Skip the governor

and

drivers/cpuidle/cpuidle.c +0 -7

-	 * always use state 0 with the tick running.
-	 */
-	if (drv->state_count <= 1) {
-		*stop_tick = false;
-		return 0;
-	}
-

  return cpuidle_curr_governor->select(drv, dev, stop_tick); }

kernel/sched/idle.c +0 -0

@@ -221,7 +221,7 @@ static void cpuidle_idle_call(void)

  next_state = cpuidle_find_deepest_state(drv, dev, max_latency_ns);
  call_cpuidle(drv, dev, next_state);

kernel/sched/idle.c +1 -1

-	} else {
+	} else if (drv->state_count > 1) {

  bool stop_tick = true;

  /*

kernel/sched/idle.c +0 -0

@@ -239,6 +239,15 @@ static void cpuidle_idle_call(void)
  		 * Give the governor an opportunity to reflect on

the outcome
  */
  cpuidle_reflect(dev, entered_state);

kernel/sched/idle.c +5 -0

+	} else {
+		tick_nohz_idle_retain_tick();
+
+		/*
+		 * If there is only a single idle state (or none),

there is

kernel/sched/idle.c +1 -0

+		 * nothing meaningful for the governor to choose. 

Skip the

kernel/sched/idle.c +3 -0

+		 * governor and always use state 0.
+		 */
+		call_cpuidle(drv, dev, 0);

  }

 exit_idle:

From: Rafael J. Wysocki <[email protected]>

If the cpuidle governor .select() callback is skipped because there is only one idle state in the cpuidle driver, the .reflect() callback should be skipped as well, at least for consistency (if not for correctness), so do it.

Fixes e5c9ffc6ae1b ("cpuidle: Skip governor when only one idle state is available")

Signed-off-by Rafael J. Wysocki <[email protected]>

Guests fall back to default_idle_call() as there is no cpuidle driver available to them by default. That causes a problem in fully loaded scenarios where CPUs go briefly idle for a couple of microseconds:

tick_nohz_idle_stop_tick() is invoked unconditionally which means unless there is timer pending in the next tick, the tick is stopped and a couple of microseconds later when the idle condition goes away restarted. That requires to program the clockevent device twice which implies a VM exit for each reprogramming.

It was suggested to remove the tick_nohz_idle_stop_tick() invocation from the default idle code, but would be counterproductive. It would not allow the host to go into deeper idle states when the guest CPU is fully idle as it has to maintain the periodic tick.

Cure this by implementing a trivial moving average filter which keeps track of the recent idle recidency time and only stop the tick when the average is larger than a tick.

Signed-off-by Thomas Gleixner <[email protected]>

Quoted 15 lines · Thomas Gleixner

Guests fall back to default_idle_call() as there is no cpuidle driver available to them by default. That causes a problem in fully loaded scenarios where CPUs go briefly idle for a couple of microseconds:

tick_nohz_idle_stop_tick() is invoked unconditionally which means unless there is timer pending in the next tick, the tick is stopped and a couple of microseconds later when the idle condition goes away restarted. That requires to program the clockevent device twice which implies a VM exit for each reprogramming.

It was suggested to remove the tick_nohz_idle_stop_tick() invocation from the default idle code, but would be counterproductive. It would not allow the host to go into deeper idle states when the guest CPU is fully idle as it has to maintain the periodic tick.

Quoted 5 lines

Cure this by implementing a trivial moving average filter which keeps track of the recent idle recidency time and only stop the tick when the average is larger than a tick.

Signed-off-by Thomas Gleixner <[email protected]>

How about so? No reason to not also pass this into the idle governors. This way it becomes a common least functionality. Governor can override, but it had better have a good reason.

---

kernel/sched/idle.c +62 -9

@@ -6,10 +6,12 @@
  * (NOTE: these are not related to SCHED_IDLE batch scheduled
  *        tasks which are handled in sched/fair.c )
  */
+#include <linux/sched/clock.h>
 #include <linux/cpuidle.h>
 #include <linux/suspend.h>
 #include <linux/livepatch.h>
 #include "sched.h"
+#include "pelt.h"
 #include "smp.h"
 
 /* Linker adds these: start and end of __cpuidle functions */

@@ -105,12 +107,7 @@ static inline void cond_tick_broadcast_e
 static inline void cond_tick_broadcast_exit(void) { }
 #endif /* !CONFIG_GENERIC_CLOCKEVENTS_BROADCAST_IDLE */
 
-/**
- * default_idle_call - Default CPU idle routine.
- *
- * To use when the cpuidle framework cannot be used.
- */
-static void __cpuidle default_idle_call(void)
+static void __cpuidle __default_idle_call(void)
 {
 	instrumentation_begin();
 	if (!current_clr_polling_and_test()) {

@@ -130,6 +127,63 @@ static void __cpuidle default_idle_call(
 	instrumentation_end();
 }
 
+#ifdef CONFIG_NO_HZ_COMMON
+
+/* Limit to 4 entries so it fits in a cache line */
+#define IDLE_DUR_ENTRIES	4
+#define IDLE_DUR_MASK		(IDLE_DUR_ENTRIES - 1)
+
+struct idle_nohz_data {
+	u64		duration[IDLE_DUR_ENTRIES];
+	u64		entry_time;
+	u64		sum;
+	unsigned int	idx;
+};
+
+static DEFINE_PER_CPU_ALIGNED(struct idle_nohz_data, nohz_data);
+
+static void default_idle_enter(void)
+{
+	this_cpu_write(nohz_data.entry_time, sched_clock());
+}
+
+static inline bool default_stop_tick(void)
+{
+	struct idle_nohz_data *nd = this_cpu_ptr(&nohz_data);
+	return nd->sum > TICK_NSEC * IDLE_DUR_ENTRIES;
+}
+
+static void default_reflect(void)
+{
+	struct idle_nohz_data *nd = this_cpu_ptr(&nohz_data);
+	unsigned int idx = nd->idx;
+	s64 delta;
+
+	/*
+	 * Build a moving average of the time spent in idle to prevent stopping
+	 * the tick on a loaded system which only goes idle briefly.
+	 */
+	delta = max(sched_clock() - nd->entry_time, 0);
+	nd->sum += delta - nd->duration[idx];
+	nd->duration[idx] = delta;
+	nd->idx = (idx + 1) & IDLE_DUR_MASK;
+}
+#else  /* CONFIG_NO_HZ_COMMON */
+static inline void default_idle_enter(void) { }
+static inline bool default_stop_tick(void) { return false; }
+static inline void default_reflect(void) { }
+#endif /* !CONFIG_NO_HZ_COMMON */
+
+static inline void default_idle_call(void)
+{
+	if (default_stop_tick())
+		tick_nohz_idle_stop_tick();
+
+	__default_idle_call();
+
+	default_reflect();
+}
+
 static int call_cpuidle_s2idle(struct cpuidle_driver *drv,
 			       struct cpuidle_device *dev,
 			       u64 max_latency_ns)

@@ -186,8 +240,6 @@ static void cpuidle_idle_call(void)
 	}
 
 	if (cpuidle_not_available(drv, dev)) {
-		tick_nohz_idle_stop_tick();
-
 		default_idle_call();
 		goto exit_idle;
 	}

@@ -222,7 +274,7 @@ static void cpuidle_idle_call(void)
 		next_state = cpuidle_find_deepest_state(drv, dev, max_latency_ns);
 		call_cpuidle(drv, dev, next_state);
 	} else {
-		bool stop_tick = true;
+		bool stop_tick = default_stop_tick();
 
 		/*
 		 * Ask the cpuidle framework to choose a convenient idle state.

@@ -238,6 +290,7 @@ static void cpuidle_idle_call(void)
 		/*
 		 * Give the governor an opportunity to reflect on the outcome
 		 */
+		default_reflect();
 		cpuidle_reflect(dev, entered_state);
 	}
 

Quoted 20 lines · Peter Zijlstra

Quoted 19 lines · Thomas Gleixner

Guests fall back to default_idle_call() as there is no cpuidle driver available to them by default. That causes a problem in fully loaded scenarios where CPUs go briefly idle for a couple of microseconds:

tick_nohz_idle_stop_tick() is invoked unconditionally which means unless there is timer pending in the next tick, the tick is stopped and a couple of microseconds later when the idle condition goes away restarted. That requires to program the clockevent device twice which implies a VM exit for each reprogramming.

It was suggested to remove the tick_nohz_idle_stop_tick() invocation from the default idle code, but would be counterproductive. It would not allow the host to go into deeper idle states when the guest CPU is fully idle as it has to maintain the periodic tick.

Cure this by implementing a trivial moving average filter which keeps track of the recent idle recidency time and only stop the tick when the average is larger than a tick.

Quoted 5 lines

Quoted 1 line

Signed-off-by Thomas Gleixner <[email protected]>

How about so? No reason to not also pass this into the idle governors. This way it becomes a common least functionality. Governor can override, but it had better have a good reason.

Quoted 67 lines

---

kernel/sched/idle.c +36 -6

@@ -6,10 +6,12 @@
  * (NOTE: these are not related to SCHED_IDLE batch scheduled
  *        tasks which are handled in sched/fair.c )
  */
+#include <linux/sched/clock.h>
 #include <linux/cpuidle.h>
 #include <linux/suspend.h>
 #include <linux/livepatch.h>
 #include "sched.h"
+#include "pelt.h"
 #include "smp.h"
 
 /* Linker adds these: start and end of __cpuidle functions */

@@ -105,12 +107,7 @@ static inline void cond_tick_broadcast_e
 static inline void cond_tick_broadcast_exit(void) { }
 #endif /* !CONFIG_GENERIC_CLOCKEVENTS_BROADCAST_IDLE */
 
-/**
- * default_idle_call - Default CPU idle routine.
- *
- * To use when the cpuidle framework cannot be used.
- */
-static void __cpuidle default_idle_call(void)
+static void __cpuidle __default_idle_call(void)
 {
      instrumentation_begin();
      if (!current_clr_polling_and_test()) {

@@ -130,6 +127,63 @@ static void __cpuidle default_idle_call(
      instrumentation_end();
 }
 
+#ifdef CONFIG_NO_HZ_COMMON
+
+/* Limit to 4 entries so it fits in a cache line */
+#define IDLE_DUR_ENTRIES     4
+#define IDLE_DUR_MASK                (IDLE_DUR_ENTRIES - 1)
+
+struct idle_nohz_data {
+     u64             duration[IDLE_DUR_ENTRIES];
+     u64             entry_time;
+     u64             sum;
+     unsigned int    idx;
+};
+
+static DEFINE_PER_CPU_ALIGNED(struct idle_nohz_data, nohz_data);
+
+static void default_idle_enter(void)
+{
+     this_cpu_write(nohz_data.entry_time, sched_clock());
+}
+
+static inline bool default_stop_tick(void)
+{
+     struct idle_nohz_data *nd = this_cpu_ptr(&nohz_data);
+     return nd->sum > TICK_NSEC * IDLE_DUR_ENTRIES;
+}
+
+static void default_reflect(void)
+{
+     struct idle_nohz_data *nd = this_cpu_ptr(&nohz_data);
+     unsigned int idx = nd->idx;
+     s64 delta;
+
+     /*

kernel/sched/idle.c static void __cpuidle default_idle_call( · line 170

Quoted 8 lines

kernel/sched/idle.c +8 -0

+      * Build a moving average of the time spent in idle to prevent stopping
+      * the tick on a loaded system which only goes idle briefly.
+      */
+     delta = max(sched_clock() - nd->entry_time, 0);
+     nd->sum += delta - nd->duration[idx];
+     nd->duration[idx] = delta;
+     nd->idx = (idx + 1) & IDLE_DUR_MASK;
+}

Quoted 46 lines

kernel/sched/idle.c +18 -3

+#else  /* CONFIG_NO_HZ_COMMON */
+static inline void default_idle_enter(void) { }
+static inline bool default_stop_tick(void) { return false; }
+static inline void default_reflect(void) { }
+#endif /* !CONFIG_NO_HZ_COMMON */
+
+static inline void default_idle_call(void)
+{
+     if (default_stop_tick())
+             tick_nohz_idle_stop_tick();
+
+     __default_idle_call();
+
+     default_reflect();
+}
+
 static int call_cpuidle_s2idle(struct cpuidle_driver *drv,
                             struct cpuidle_device *dev,
                             u64 max_latency_ns)

@@ -186,8 +240,6 @@ static void cpuidle_idle_call(void)
      }
 
      if (cpuidle_not_available(drv, dev)) {
-             tick_nohz_idle_stop_tick();
-
              default_idle_call();
              goto exit_idle;
      }

@@ -222,7 +274,7 @@ static void cpuidle_idle_call(void)
              next_state = cpuidle_find_deepest_state(drv, dev, max_latency_ns);
              call_cpuidle(drv, dev, next_state);
      } else {
-             bool stop_tick = true;
+             bool stop_tick = default_stop_tick();
 
              /*
               * Ask the cpuidle framework to choose a convenient idle state.

@@ -238,6 +290,7 @@ static void cpuidle_idle_call(void)
              /*
               * Give the governor an opportunity to reflect on the outcome
               */
+             default_reflect();
              cpuidle_reflect(dev, entered_state);
      }
 

kernel/sched/idle.c static void do_idle(void) · line 335

Quoted 7 lines

kernel/sched/idle.c +1 -0

@@ -276,6 +329,7 @@ static void do_idle(void)
 
      __current_set_polling();
      tick_nohz_idle_enter();
+     default_idle_enter();
 
      while (!need_resched()) {
 

Damn, lost hunk:

drivers/cpuidle/cpuidle.c +0 -10

index 65fbb8e807b9..c7876e9e024f 100644
@@ -359,16 +359,6 @@ noinstr int cpuidle_enter_state(struct cpuidle_device *dev,
 int cpuidle_select(struct cpuidle_driver *drv, struct cpuidle_device *dev,
                   bool *stop_tick)
 {
-       /*
-        * If there is only a single idle state (or none), there is nothing
-        * meaningful for the governor to choose. Skip the governor and
-        * always use state 0 with the tick running.
-        */
-       if (drv->state_count <= 1) {
-               *stop_tick = false;
-               return 0;
-       }
-
        return cpuidle_curr_governor->select(drv, dev, stop_tick);
 }
 

Also, I suppose menu wants this?

drivers/cpuidle/governors/menu.c +2 -1

index 899ff16ff1fe..a75fe1fca65d 100644
@@ -290,7 +290,8 @@ static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev,
                 * it right away and keep the tick running if state[0] is a
                 * polling one.
                 */
-               *stop_tick = !(drv->states[0].flags & CPUIDLE_FLAG_POLLING);
+               if (drv->states[0].flags & CPUIDLE_FLAG_POLLING)
+                       *stop_tick = false;
                return 0;
        }