cpuidle: governors: menu: A fix, a corner case adjustment and a cleanup

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

Occasionally, the exit latency of the idle state selected by the menu governor may exceed the PM QoS CPU wakeup latency limit. Namely, if the scheduler tick has been stopped already and predicted_ns is greater than the tick period length, the governor may return an idle state whose exit latency exceeds latency_req because that decision is made before checking the current idle state's exit latency.

For instance, say that there are 3 idle states, 0, 1, and 2. For idle states 0 and 1, the exit latency is equal to the target residency and the values are 0 and 5 us, respectively. State 2 is deeper and has the exit latency and target residency of 200 us and 2 ms (which is greater than the tick period length), respectively.

Say that predicted_ns is equal to TICK_NSEC and the PM QoS latency limit is 20 us. After the first two iterations of the main loop in menu_select(), idx becomes 1 and in the third iteration of it the target

residency of the current state (state 2) is greater than predicted_ns. State 2 is not a polling one and predicted_ns is not less than TICK_NSEC, so the check on whether or not the tick has been stopped is done. Say that the tick has been stopped already and there are no imminent timers (that is, delta_tick is greater than the target residency of state 2). In that case, idx becomes 2 and it is returned immediately, but the exit latency of state 2 exceeds the latency limit.

Address this issue by modifying the code to compare the exit latency of the current idle state (idle state i) with the latency limit before comparing its target residecy with predicted_ns, which allows one

more exit_latency_ns check that becomes redundant to be dropped.

However, after the above change, latency_req cannot take the predicted_ns value any more, which takes place after commit 38f83090f515 ("cpuidle:
menu: Remove iowait influence"), because it may cause a polling state to be returned prematurely.

In the context of the previous example say that predicted_ns is 3000 and the PM QoS latency limit is still 20 us. Additionally, say that idle state 0 is a polling one. Moving the exit_latency_ns check before the target_residency_ns one causes the loop to terminate in the second iteration, before the target_residency_ns check, so idle state 0 will be returned even though previously state 1 would be returned if there were no imminent timers.

For this reason, remove the assignment of the predicted_ns value to latency_req from the code.

Fixes 5ef499cd571c ("cpuidle: menu: Handle stopped tick more aggressively")

Cc 4.17+ <[email protected]> # 4.17+

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

---
 drivers/cpuidle/governors/menu.c |   29 ++++++++++++-----------------
 1 file changed, 12 insertions(+), 17 deletions(-)

drivers/cpuidle/governors/menu.c +12 -15

@@ -287,20 +287,15 @@
 		return 0;
 	}
 
-	if (tick_nohz_tick_stopped()) {
-		/*
-		 * If the tick is already stopped, the cost of possible short
-		 * idle duration misprediction is much higher, because the CPU
-		 * may be stuck in a shallow idle state for a long time as a
-		 * result of it.  In that case say we might mispredict and use
-		 * the known time till the closest timer event for the idle
-		 * state selection.
-		 */
-		if (predicted_ns < TICK_NSEC)
-			predicted_ns = data->next_timer_ns;
-	} else if (latency_req > predicted_ns) {
-		latency_req = predicted_ns;
-	}
+	/*
+	 * If the tick is already stopped, the cost of possible short idle
+	 * duration misprediction is much higher, because the CPU may be stuck
+	 * in a shallow idle state for a long time as a result of it.  In that
+	 * case, say we might mispredict and use the known time till the closest
+	 * timer event for the idle state selection.
+	 */
+	if (tick_nohz_tick_stopped() && predicted_ns < TICK_NSEC)
+		predicted_ns = data->next_timer_ns;
 
 	/*
 	 * Find the idle state with the lowest power while satisfying

@@ -316,13 +311,15 @@
 		if (idx == -1)
 			idx = i; /* first enabled state */
 
+		if (s->exit_latency_ns > latency_req)
+			break;
+
 		if (s->target_residency_ns > predicted_ns) {
 			/*
 			 * Use a physical idle state, not busy polling, unless
 			 * a timer is going to trigger soon enough.
 			 */
 			if ((drv->states[idx].flags & CPUIDLE_FLAG_POLLING) &&
-			    s->exit_latency_ns <= latency_req &&
 			    s->target_residency_ns <= data->next_timer_ns) {
 				predicted_ns = s->target_residency_ns;
 				idx = i;

Quoted 15 lines · Rafael J. Wysocki

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

Occasionally, the exit latency of the idle state selected by the menu governor may exceed the PM QoS CPU wakeup latency limit. Namely, if the scheduler tick has been stopped already and predicted_ns is greater than the tick period length, the governor may return an idle state whose exit latency exceeds latency_req because that decision is made before checking the current idle state's exit latency.

For instance, say that there are 3 idle states, 0, 1, and 2. For idle states 0 and 1, the exit latency is equal to the target residency and the values are 0 and 5 us, respectively. State 2 is deeper and has the exit latency and target residency of 200 us and 2 ms (which is greater than the tick period length), respectively.

Quoted 3 lines

Say that predicted_ns is equal to TICK_NSEC and the PM QoS latency limit is 20 us. After the first two iterations of the main loop in menu_select(), idx becomes 1 and in the third iteration of it the target

Can drop "of it" here?

Quoted 8 lines

residency of the current state (state 2) is greater than predicted_ns. State 2 is not a polling one and predicted_ns is not less than TICK_NSEC, so the check on whether or not the tick has been stopped is done. Say that the tick has been stopped already and there are no imminent timers (that is, delta_tick is greater than the target residency of state 2). In that case, idx becomes 2 and it is returned immediately, but the exit latency of state 2 exceeds the latency limit.

Quoted 3 lines

Address this issue by modifying the code to compare the exit latency of the current idle state (idle state i) with the latency limit before comparing its target residecy with predicted_ns, which allows one

residency

Quoted 81 lines

more exit_latency_ns check that becomes redundant to be dropped.

However, after the above change, latency_req cannot take the predicted_ns value any more, which takes place after commit 38f83090f515 ("cpuidle:
menu: Remove iowait influence"), because it may cause a polling state to be returned prematurely.

In the context of the previous example say that predicted_ns is 3000 and the PM QoS latency limit is still 20 us. Additionally, say that idle state 0 is a polling one. Moving the exit_latency_ns check before the target_residency_ns one causes the loop to terminate in the second iteration, before the target_residency_ns check, so idle state 0 will be returned even though previously state 1 would be returned if there were no imminent timers.

For this reason, remove the assignment of the predicted_ns value to latency_req from the code.

Fixes 5ef499cd571c ("cpuidle: menu: Handle stopped tick more aggressively")

Cc 4.17+ <[email protected]> # 4.17+

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

---
 drivers/cpuidle/governors/menu.c |   29 ++++++++++++-----------------
 1 file changed, 12 insertions(+), 17 deletions(-)

drivers/cpuidle/governors/menu.c +12 -17

@@ -287,20 +287,15 @@
              return 0;
      }
 
-     if (tick_nohz_tick_stopped()) {
-             /*
-              * If the tick is already stopped, the cost of possible short
-              * idle duration misprediction is much higher, because the CPU
-              * may be stuck in a shallow idle state for a long time as a
-              * result of it.  In that case say we might mispredict and use
-              * the known time till the closest timer event for the idle
-              * state selection.
-              */
-             if (predicted_ns < TICK_NSEC)
-                     predicted_ns = data->next_timer_ns;
-     } else if (latency_req > predicted_ns) {
-             latency_req = predicted_ns;
-     }
+     /*
+      * If the tick is already stopped, the cost of possible short idle
+      * duration misprediction is much higher, because the CPU may be stuck
+      * in a shallow idle state for a long time as a result of it.  In that
+      * case, say we might mispredict and use the known time till the closest
+      * timer event for the idle state selection.
+      */
+     if (tick_nohz_tick_stopped() && predicted_ns < TICK_NSEC)
+             predicted_ns = data->next_timer_ns;
 
      /*
       * Find the idle state with the lowest power while satisfying

@@ -316,13 +311,15 @@
              if (idx == -1)
                      idx = i; /* first enabled state */
 
+             if (s->exit_latency_ns > latency_req)
+                     break;
+
              if (s->target_residency_ns > predicted_ns) {
                      /*
                       * Use a physical idle state, not busy polling, unless
                       * a timer is going to trigger soon enough.
                       */
                      if ((drv->states[idx].flags & CPUIDLE_FLAG_POLLING) &&
-                         s->exit_latency_ns <= latency_req &&
                          s->target_residency_ns <= data->next_timer_ns) {
                              predicted_ns = s->target_residency_ns;
                              idx = i;

@@ -354,8 +351,6 @@
 
                      return idx;
              }
-             if (s->exit_latency_ns > latency_req)
-                     break;
 

drivers/cpuidle/governors/menu.c line 356

Quoted 5 lines

drivers/cpuidle/governors/menu.c +0 -0

              idx = i;
      }

Good catch!

Reviewed-by Christian Loehle <[email protected]>

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

Occasionally, the exit latency of the idle state selected by the menu governor may exceed the PM QoS CPU wakeup latency limit. Namely, if the scheduler tick has been stopped already and predicted_ns is greater than the tick period length, the governor may return an idle state whose exit latency exceeds latency_req because that decision is made before checking the current idle state's exit latency.

For instance, say that there are 3 idle states, 0, 1, and 2. For idle states 0 and 1, the exit latency is equal to the target residency and the values are 0 and 5 us, respectively. State 2 is deeper and has the exit latency and target residency of 200 us and 2 ms (which is greater than the tick period length), respectively.

Say that predicted_ns is equal to TICK_NSEC and the PM QoS latency limit is 20 us. After the first two iterations of the main loop in menu_select(), idx becomes 1 and in the third iteration of it the target residency of the current state (state 2) is greater than predicted_ns. State 2 is not a polling one and predicted_ns is not less than TICK_NSEC, so the check on whether or not the tick has been stopped is done. Say that the tick has been stopped already and there are no imminent timers (that is, delta_tick is greater than the target residency of state 2). In that case, idx becomes 2 and it is returned immediately, but the exit latency of state 2 exceeds the latency limit.

Address this issue by modifying the code to compare the exit latency of the current idle state (idle state i) with the latency limit before comparing its target residecy with predicted_ns, which allows one more exit_latency_ns check that becomes redundant to be dropped.

However, after the above change, latency_req cannot take the predicted_ns value any more, which takes place after commit 38f83090f515 ("cpuidle:
menu: Remove iowait influence"), because it may cause a polling state to be returned prematurely.

In the context of the previous example say that predicted_ns is 3000 and the PM QoS latency limit is still 20 us. Additionally, say that idle state 0 is a polling one. Moving the exit_latency_ns check before the target_residency_ns one causes the loop to terminate in the second iteration, before the target_residency_ns check, so idle state 0 will be returned even though previously state 1 would be returned if there were no imminent timers.

For this reason, remove the assignment of the predicted_ns value to latency_req from the code.

Fixes 5ef499cd571c ("cpuidle: menu: Handle stopped tick more aggressively")

Cc 4.17+ <[email protected]> # 4.17+

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

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

However, after the above change, latency_req cannot take the predicted_ns value any more, which takes place after commit 38f83090f515 ("cpuidle:
menu: Remove iowait influence"), because it may cause a polling state to be returned prematurely.

In the context of the previous example say that predicted_ns is 3000 and the PM QoS latency limit is still 20 us. Additionally, say that idle state 0 is a polling one. Moving the exit_latency_ns check before the target_residency_ns one causes the loop to terminate in the second iteration, before the target_residency_ns check, so idle state 0 will be returned even though previously state 1 would be returned if there were no imminent timers.

For this reason, remove the assignment of the predicted_ns value to latency_req from the code.

Hi Rafael,

Recent email communications about other patches had me looking at this one again.

Quoted 2 lines · On 2025.08.13 03:26 Rafael wrote:

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

... snip...

Quoted 13 lines

However, after the above change, latency_req cannot take the predicted_ns value any more, which takes place after commit 38f83090f515 ("cpuidle:
menu: Remove iowait influence"), because it may cause a polling state to be returned prematurely.

In the context of the previous example say that predicted_ns is 3000 and the PM QoS latency limit is still 20 us. Additionally, say that idle state 0 is a polling one. Moving the exit_latency_ns check before the target_residency_ns one causes the loop to terminate in the second iteration, before the target_residency_ns check, so idle state 0 will be returned even though previously state 1 would be returned if there were no imminent timers.

Quoted 2 lines

For this reason, remove the assignment of the predicted_ns value to latency_req from the code.

Which is okay for timer-based workflow,
but what about non-timer based, or interrupt driven, workflow?

Under conditions where idle state 0, or Polling, would be used a lot, I am observing about a 11 % throughput regression with this patch And idle state 0, polling, usage going from 20% to 0%.

From my testing of kernels 6.17-rc1, rc2,rc3 in August and September and again now. I missed this in August/September:

779b1a1cb13a cpuidle: governors: menu: Avoid selecting states with too much latency - v6.17-rc3 fa3fa55de0d6 cpuidle: governors: menu: Avoid using invalid recent intervals data - v6.17-rc2 baseline reference: v6.17-rc1

teo was included also. As far as I can recall its response has always been similar to rc3. At least, recently.

Three graphs are attached:
Sampling data once per 20 seconds, the test is started after the first idle sample, and at least one sample is taken after the system returns to idle after the test. The faster the test runs the better.

Test computer:
Processor: Intel(R) Core(TM) i5-10600K CPU @ 4.10GHz
Distro: Ubuntu 24.04.3, server, no desktop GUI. CPU frequency scaling driver: intel_pstate
HWP: disabled.
CPU frequency scaling governor: performance
Ilde driver: intel_idle
Idle governor: menu (except teo for one compare test run) Idle states: 4: name : description:
state0/name:POLL desc:CPUIDLE CORE POLL IDLE

  state1/name:C1_ACPI          desc:ACPI FFH MWAIT 0x0
  state2/name:C2_ACPI          desc:ACPI FFH MWAIT 0x30
  state3/name:C3_ACPI          desc:ACPI FFH MWAIT 0x60

Hi Doug,

Quoted 57 lines · CEST Doug Smythies

Hi Rafael,

Recent email communications about other patches had me looking at this one again.

Quoted 2 lines · On 2025.08.13 03:26 Rafael wrote:

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

... snip...

Quoted 13 lines

However, after the above change, latency_req cannot take the predicted_ns value any more, which takes place after commit 38f83090f515 ("cpuidle:
menu: Remove iowait influence"), because it may cause a polling state to be returned prematurely.

In the context of the previous example say that predicted_ns is 3000 and the PM QoS latency limit is still 20 us. Additionally, say that idle state 0 is a polling one. Moving the exit_latency_ns check before the target_residency_ns one causes the loop to terminate in the second iteration, before the target_residency_ns check, so idle state 0 will be returned even though previously state 1 would be returned if there were no imminent timers.

Quoted 2 lines

For this reason, remove the assignment of the predicted_ns value to latency_req from the code.

Which is okay for timer-based workflow,
but what about non-timer based, or interrupt driven, workflow?

Under conditions where idle state 0, or Polling, would be used a lot, I am observing about a 11 % throughput regression with this patch And idle state 0, polling, usage going from 20% to 0%.

From my testing of kernels 6.17-rc1, rc2,rc3 in August and September and again now. I missed this in August/September:

779b1a1cb13a cpuidle: governors: menu: Avoid selecting states with too much latency - v6.17-rc3 fa3fa55de0d6 cpuidle: governors: menu: Avoid using invalid recent intervals data - v6.17-rc2 baseline reference: v6.17-rc1

teo was included also. As far as I can recall its response has always been similar to rc3. At least, recently.

Three graphs are attached:
Sampling data once per 20 seconds, the test is started after the first idle sample, and at least one sample is taken after the system returns to idle after the test. The faster the test runs the better.

Test computer:
Processor: Intel(R) Core(TM) i5-10600K CPU @ 4.10GHz
Distro: Ubuntu 24.04.3, server, no desktop GUI. CPU frequency scaling driver: intel_pstate
HWP: disabled.
CPU frequency scaling governor: performance
Ilde driver: intel_idle
Idle governor: menu (except teo for one compare test run) Idle states: 4: name : description:
state0/name:POLL desc:CPUIDLE CORE POLL IDLE

Quoted 3 lines

  state1/name:C1_ACPI          desc:ACPI FFH MWAIT 0x0
  state2/name:C2_ACPI          desc:ACPI FFH MWAIT 0x30
  state3/name:C3_ACPI          desc:ACPI FFH MWAIT 0x60

OK, so since the exit residency of an idle state cannot exceed its target residency, the appended change (on top of 6.18-rc2) should make the throughput regression go away.

---
drivers/cpuidle/governors/menu.c | 7 +++++--

1 file changed, 5 insertions(+), 2 deletions(-)

drivers/cpuidle/governors/menu.c +5 -2

@@ -321,10 +321,13 @@ static int menu_select(struct cpuidle_dr
 
 		/*
 		 * Use a physical idle state, not busy polling, unless a timer
-		 * is going to trigger soon enough.
+		 * is going to trigger soon enough or the exit latency of the
+		 * idle state in question is greater than the predicted idle
+		 * duration.
 		 */
 		if ((drv->states[idx].flags & CPUIDLE_FLAG_POLLING) &&
-		    s->target_residency_ns <= data->next_timer_ns) {
+		    s->target_residency_ns <= data->next_timer_ns &&
+		    s->exit_latency_ns <= predicted_ns) {
 			predicted_ns = s->target_residency_ns;
 			idx = i;
 			break;

Quoted 64 lines · On 2025.10.23 07:51 Rafael wrote:

Hi Doug,

Quoted 60 lines · CEST Doug Smythies

Hi Rafael,

Recent email communications about other patches had me looking at this one again.

Quoted 2 lines · On 2025.08.13 03:26 Rafael wrote:

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

... snip...

Quoted 13 lines

However, after the above change, latency_req cannot take the predicted_ns value any more, which takes place after commit 38f83090f515 ("cpuidle:
menu: Remove iowait influence"), because it may cause a polling state to be returned prematurely.

In the context of the previous example say that predicted_ns is 3000 and the PM QoS latency limit is still 20 us. Additionally, say that idle state 0 is a polling one. Moving the exit_latency_ns check before the target_residency_ns one causes the loop to terminate in the second iteration, before the target_residency_ns check, so idle state 0 will be returned even though previously state 1 would be returned if there were no imminent timers.

Quoted 2 lines

For this reason, remove the assignment of the predicted_ns value to latency_req from the code.

Which is okay for timer-based workflow,
but what about non-timer based, or interrupt driven, workflow?

Under conditions where idle state 0, or Polling, would be used a lot, I am observing about a 11 % throughput regression with this patch And idle state 0, polling, usage going from 20% to 0%.

From my testing of kernels 6.17-rc1, rc2,rc3 in August and September and again now. I missed this in August/September:

779b1a1cb13a cpuidle: governors: menu: Avoid selecting states with too much latency - v6.17-rc3 fa3fa55de0d6 cpuidle: governors: menu: Avoid using invalid recent intervals data - v6.17-rc2 baseline reference: v6.17-rc1

teo was included also. As far as I can recall its response has always been similar to rc3. At least, recently.

Three graphs are attached:
Sampling data once per 20 seconds, the test is started after the first idle sample, and at least one sample is taken after the system returns to idle after the test. The faster the test runs the better.

Test computer:
Processor: Intel(R) Core(TM) i5-10600K CPU @ 4.10GHz
Distro: Ubuntu 24.04.3, server, no desktop GUI. CPU frequency scaling driver: intel_pstate
HWP: disabled.
CPU frequency scaling governor: performance
Ilde driver: intel_idle
Idle governor: menu (except teo for one compare test run) Idle states: 4: name : description:
state0/name:POLL desc:CPUIDLE CORE POLL IDLE

  state1/name:C1_ACPI          desc:ACPI FFH MWAIT 0x0
  state2/name:C2_ACPI          desc:ACPI FFH MWAIT 0x30
  state3/name:C3_ACPI          desc:ACPI FFH MWAIT 0x60

Quoted 3 lines

OK, so since the exit residency of an idle state cannot exceed its target residency, the appended change (on top of 6.18-rc2) should make the throughput regression go away.

Indeed, the patch you appended below did make the
throughput regression go away.

Thank you.

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

Reduce the indentation level in the main loop of menu_select() by rearranging some checks and assignments in it.

No intentional functional impact.

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

---
 drivers/cpuidle/governors/menu.c |   72 ++++++++++++++++++++-------------------
 1 file changed, 37 insertions(+), 35 deletions(-)

drivers/cpuidle/governors/menu.c +36 -34

@@ -314,45 +314,47 @@
 		if (s->exit_latency_ns > latency_req)
 			break;
 
-		if (s->target_residency_ns > predicted_ns) {
-			/*
-			 * Use a physical idle state, not busy polling, unless
-			 * a timer is going to trigger soon enough.
-			 */
-			if ((drv->states[idx].flags & CPUIDLE_FLAG_POLLING) &&
-			    s->target_residency_ns <= data->next_timer_ns) {
-				predicted_ns = s->target_residency_ns;
-				idx = i;
-				break;
-			}
-			if (predicted_ns < TICK_NSEC)
-				break;
-
-			if (!tick_nohz_tick_stopped()) {
-				/*
-				 * If the state selected so far is shallow,
-				 * waking up early won't hurt, so retain the
-				 * tick in that case and let the governor run
-				 * again in the next iteration of the loop.
-				 */
-				predicted_ns = drv->states[idx].target_residency_ns;
-				break;
-			}
+		if (s->target_residency_ns <= predicted_ns) {
+			idx = i;
+			continue;
+		}
+
+		/*
+		 * Use a physical idle state, not busy polling, unless a timer
+		 * is going to trigger soon enough.
+		 */
+		if ((drv->states[idx].flags & CPUIDLE_FLAG_POLLING) &&
+		    s->target_residency_ns <= data->next_timer_ns) {
+			predicted_ns = s->target_residency_ns;
+			idx = i;
+			break;
+		}
 
+		if (predicted_ns < TICK_NSEC)
+			break;
+
+		if (!tick_nohz_tick_stopped()) {
 			/*
-			 * If the state selected so far is shallow and this
-			 * state's target residency matches the time till the
-			 * closest timer event, select this one to avoid getting
-			 * stuck in the shallow one for too long.
+			 * If the state selected so far is shallow, waking up
+			 * early won't hurt, so retain the tick in that case and
+			 * let the governor run again in the next iteration of
+			 * the idle loop.
 			 */
-			if (drv->states[idx].target_residency_ns < TICK_NSEC &&
-			    s->target_residency_ns <= delta_tick)
-				idx = i;
-
-			return idx;
+			predicted_ns = drv->states[idx].target_residency_ns;
+			break;
 		}
 
-		idx = i;
+		/*
+		 * If the state selected so far is shallow and this state's
+		 * target residency matches the time till the closest timer
+		 * event, select this one to avoid getting stuck in the shallow
+		 * one for too long.
+		 */
+		if (drv->states[idx].target_residency_ns < TICK_NSEC &&
+		    s->target_residency_ns <= delta_tick)
+			idx = i;
+
+		return idx;
 	}
 
 	if (idx == -1)

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

Reduce the indentation level in the main loop of menu_select() by rearranging some checks and assignments in it.

No intentional functional impact.

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

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

When the menu governor runs on a nohz_full CPU and there are no user space timers in the workload on that CPU, it ends up selecting idle states with target residency values above TICK_NSEC all the time due to a tick_nohz_tick_stopped() check designed for a different use case. Namely, on nohz_full CPUs the fact that the tick has been stopped does not actually mean anything in particular, whereas in the other case it indicates that previously the CPU was expected to be idle sufficiently long for the tick to be stopped, so it is not unreasonable to expect it to be idle beyond the tick period length again.

In some cases, this behavior causes latency in the workload to grow undesirably. It may also cause the workload to consume more energy than necessary if the CPU does not spend enough time in the selected deep idle states.

Address this by amending the tick_nohz_tick_stopped() check in question with a tick_nohz_full_cpu() one to avoid using the time till the next timer event as the predicted_ns value all the time on nohz_full CPUs.

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

---
 drivers/cpuidle/governors/menu.c |   12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)

drivers/cpuidle/governors/menu.c +6 -0

@@ -293,8 +293,18 @@
 	 * in a shallow idle state for a long time as a result of it.  In that
 	 * case, say we might mispredict and use the known time till the closest
 	 * timer event for the idle state selection.
+	 *
+	 * However, on nohz_full CPUs the tick does not run as a rule and the
+	 * time till the closest timer event may always be effectively infinite,
+	 * so using it as a replacement for the predicted idle duration would
+	 * effectively always cause the prediction results to be discarded and
+	 * deep idle states to be selected all the time.  That might introduce

Quoted 47 lines · Rafael J. Wysocki

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

When the menu governor runs on a nohz_full CPU and there are no user space timers in the workload on that CPU, it ends up selecting idle states with target residency values above TICK_NSEC all the time due to a tick_nohz_tick_stopped() check designed for a different use case. Namely, on nohz_full CPUs the fact that the tick has been stopped does not actually mean anything in particular, whereas in the other case it indicates that previously the CPU was expected to be idle sufficiently long for the tick to be stopped, so it is not unreasonable to expect it to be idle beyond the tick period length again.

In some cases, this behavior causes latency in the workload to grow undesirably. It may also cause the workload to consume more energy than necessary if the CPU does not spend enough time in the selected deep idle states.

Address this by amending the tick_nohz_tick_stopped() check in question with a tick_nohz_full_cpu() one to avoid using the time till the next timer event as the predicted_ns value all the time on nohz_full CPUs.

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

---
 drivers/cpuidle/governors/menu.c |   12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)

drivers/cpuidle/governors/menu.c +11 -1

@@ -293,8 +293,18 @@
       * in a shallow idle state for a long time as a result of it.  In that
       * case, say we might mispredict and use the known time till the closest
       * timer event for the idle state selection.
+      *
+      * However, on nohz_full CPUs the tick does not run as a rule and the
+      * time till the closest timer event may always be effectively infinite,
+      * so using it as a replacement for the predicted idle duration would
+      * effectively always cause the prediction results to be discarded and
+      * deep idle states to be selected all the time.  That might introduce
+      * unwanted latency into the workload and cause more energy than
+      * necessary to be consumed if the discarded prediction results are
+      * actually accurate, so skip nohz_full CPUs here.
       */
-     if (tick_nohz_tick_stopped() && predicted_ns < TICK_NSEC)
+     if (tick_nohz_tick_stopped() && !tick_nohz_full_cpu(dev->cpu) &&
+         predicted_ns < TICK_NSEC)
              predicted_ns = data->next_timer_ns;
 

drivers/cpuidle/governors/menu.c line 310

Quoted 4 lines

drivers/cpuidle/governors/menu.c +0 -0

      /*

OTOH the behaviour with $SUBJECT possibly means that we use predicted_ns from get_typical_interval() (which may suggest picking a shallow state based on previous wakeup patterns) only then to never wake up again?

Quoted 49 lines · Christian Loehle

Quoted 47 lines · Rafael J. Wysocki

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

When the menu governor runs on a nohz_full CPU and there are no user space timers in the workload on that CPU, it ends up selecting idle states with target residency values above TICK_NSEC all the time due to a tick_nohz_tick_stopped() check designed for a different use case. Namely, on nohz_full CPUs the fact that the tick has been stopped does not actually mean anything in particular, whereas in the other case it indicates that previously the CPU was expected to be idle sufficiently long for the tick to be stopped, so it is not unreasonable to expect it to be idle beyond the tick period length again.

In some cases, this behavior causes latency in the workload to grow undesirably. It may also cause the workload to consume more energy than necessary if the CPU does not spend enough time in the selected deep idle states.

Address this by amending the tick_nohz_tick_stopped() check in question with a tick_nohz_full_cpu() one to avoid using the time till the next timer event as the predicted_ns value all the time on nohz_full CPUs.

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

---
 drivers/cpuidle/governors/menu.c |   12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)

drivers/cpuidle/governors/menu.c +11 -1

@@ -293,8 +293,18 @@
       * in a shallow idle state for a long time as a result of it.  In that
       * case, say we might mispredict and use the known time till the closest
       * timer event for the idle state selection.
+      *
+      * However, on nohz_full CPUs the tick does not run as a rule and the
+      * time till the closest timer event may always be effectively infinite,
+      * so using it as a replacement for the predicted idle duration would
+      * effectively always cause the prediction results to be discarded and
+      * deep idle states to be selected all the time.  That might introduce
+      * unwanted latency into the workload and cause more energy than
+      * necessary to be consumed if the discarded prediction results are
+      * actually accurate, so skip nohz_full CPUs here.
       */
-     if (tick_nohz_tick_stopped() && predicted_ns < TICK_NSEC)
+     if (tick_nohz_tick_stopped() && !tick_nohz_full_cpu(dev->cpu) &&
+         predicted_ns < TICK_NSEC)
              predicted_ns = data->next_timer_ns;
 

Quoted 8 lines

drivers/cpuidle/governors/menu.c line 310

Quoted 4 lines

drivers/cpuidle/governors/menu.c +0 -0

      /*

OTOH the behaviour with $SUBJECT possibly means that we use predicted_ns from get_typical_interval() (which may suggest picking a shallow state based on previous wakeup patterns) only then to never wake up again?

Yes, there is this risk, but the current behavior is more damaging IMV because it (potentially) hurts both energy efficiency and performance.

It is also arguably easier for the user to remedy getting stuck in a shallow idle state than to change governor's behavior (PM QoS is a bit too blunt for this).

Moreover, configuring CPUs as nohz_full and leaving them in long idle may not be the most efficient use of them.

Quoted 67 lines · Rafael J. Wysocki

Quoted 49 lines · Christian Loehle

Quoted 47 lines · Rafael J. Wysocki

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

When the menu governor runs on a nohz_full CPU and there are no user space timers in the workload on that CPU, it ends up selecting idle states with target residency values above TICK_NSEC all the time due to a tick_nohz_tick_stopped() check designed for a different use case. Namely, on nohz_full CPUs the fact that the tick has been stopped does not actually mean anything in particular, whereas in the other case it indicates that previously the CPU was expected to be idle sufficiently long for the tick to be stopped, so it is not unreasonable to expect it to be idle beyond the tick period length again.

In some cases, this behavior causes latency in the workload to grow undesirably. It may also cause the workload to consume more energy than necessary if the CPU does not spend enough time in the selected deep idle states.

Address this by amending the tick_nohz_tick_stopped() check in question with a tick_nohz_full_cpu() one to avoid using the time till the next timer event as the predicted_ns value all the time on nohz_full CPUs.

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

---
 drivers/cpuidle/governors/menu.c |   12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)

drivers/cpuidle/governors/menu.c +11 -1

@@ -293,8 +293,18 @@
       * in a shallow idle state for a long time as a result of it.  In that
       * case, say we might mispredict and use the known time till the closest
       * timer event for the idle state selection.
+      *
+      * However, on nohz_full CPUs the tick does not run as a rule and the
+      * time till the closest timer event may always be effectively infinite,
+      * so using it as a replacement for the predicted idle duration would
+      * effectively always cause the prediction results to be discarded and
+      * deep idle states to be selected all the time.  That might introduce
+      * unwanted latency into the workload and cause more energy than
+      * necessary to be consumed if the discarded prediction results are
+      * actually accurate, so skip nohz_full CPUs here.
       */
-     if (tick_nohz_tick_stopped() && predicted_ns < TICK_NSEC)
+     if (tick_nohz_tick_stopped() && !tick_nohz_full_cpu(dev->cpu) &&
+         predicted_ns < TICK_NSEC)
              predicted_ns = data->next_timer_ns;
 

Quoted 8 lines

drivers/cpuidle/governors/menu.c line 310

Quoted 4 lines

drivers/cpuidle/governors/menu.c +0 -0

      /*

OTOH the behaviour with $SUBJECT possibly means that we use predicted_ns from get_typical_interval() (which may suggest picking a shallow state based on previous wakeup patterns) only then to never wake up again?

Yes, there is this risk, but the current behavior is more damaging IMV because it (potentially) hurts both energy efficiency and performance.

It is also arguably easier for the user to remedy getting stuck in a shallow idle state than to change governor's behavior (PM QoS is a bit too blunt for this).

Quoted 2 lines

Moreover, configuring CPUs as nohz_full and leaving them in long idle may not be the most efficient use of them.

True, on the other hand the setup cost for nohz_full is so high, you'd expect the additional idle states disabling depending on the workload isn't too much to ask for...

Anyway feel free to go ahead.

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

When the menu governor runs on a nohz_full CPU and there are no user space timers in the workload on that CPU, it ends up selecting idle states with target residency values above TICK_NSEC all the time due to a tick_nohz_tick_stopped() check designed for a different use case.

Namely, on nohz_full CPUs the fact that the tick has been stopped does not actually mean anything in particular, whereas in the other case it indicates that previously the CPU was expected to be idle sufficiently long for the tick to be stopped, so it is not unreasonable to expect it to be idle beyond the tick period length again.

In some cases, this behavior causes latency in the workload to grow undesirably. It may also cause the workload to consume more energy than necessary if the CPU does not spend enough time in the selected deep idle states.

Address this by amending the tick_nohz_tick_stopped() check in question with a tick_nohz_full_cpu() one to avoid using the time till the next timer event as the predicted_ns value all the time on nohz_full CPUs.

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

---
 drivers/cpuidle/governors/menu.c |   12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)

drivers/cpuidle/governors/menu.c +4 -0

@@ -293,8 +293,18 @@
 	 * in a shallow idle state for a long time as a result of it.  In that
 	 * case, say we might mispredict and use the known time till the closest
 	 * timer event for the idle state selection.
+	 *
+	 * However, on nohz_full CPUs the tick does not run as a rule and the
+	 * time till the closest timer event may always be effectively infinite,
+	 * so using it as a replacement for the predicted idle duration would

drivers/cpuidle/governors/menu.c +0 -0

 	/*
 

Quoted 7 lines · Rafael J. Wysocki

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

When the menu governor runs on a nohz_full CPU and there are no user space timers in the workload on that CPU, it ends up selecting idle states with target residency values above TICK_NSEC all the time due to a tick_nohz_tick_stopped() check designed for a different use case.

Quoted 5 lines

Namely, on nohz_full CPUs the fact that the tick has been stopped does not actually mean anything in particular, whereas in the other case it indicates that previously the CPU was expected to be idle sufficiently long for the tick to be stopped, so it is not unreasonable to expect it to be idle beyond the tick period length again.

I understand what you mean but it may be hard to figure out for reviewers. Can we rephrase it to something like:

When nohz_full is not running, the fact that the tick is stopped indicates the CPU has been idle for sufficiently long so that nohz has deferred it to the next timer callback. So it is not unreasonable to expect the CPU to be idle beyond the tick period length again.

However when nohz_full is running, the CPU may enter idle with the tick already stopped. But this doesn't tell anything about the future CPU's idleness.

Quoted 25 lines

In some cases, this behavior causes latency in the workload to grow undesirably. It may also cause the workload to consume more energy than necessary if the CPU does not spend enough time in the selected deep idle states.

Address this by amending the tick_nohz_tick_stopped() check in question with a tick_nohz_full_cpu() one to avoid using the time till the next timer event as the predicted_ns value all the time on nohz_full CPUs.

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

---
 drivers/cpuidle/governors/menu.c |   12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)

drivers/cpuidle/governors/menu.c +4 -0

@@ -293,8 +293,18 @@
       * in a shallow idle state for a long time as a result of it.  In that
       * case, say we might mispredict and use the known time till the closest
       * timer event for the idle state selection.
+      *
+      * However, on nohz_full CPUs the tick does not run as a rule and the
+      * time till the closest timer event may always be effectively infinite,
+      * so using it as a replacement for the predicted idle duration would

drivers/cpuidle/governors/menu.c line 308

Quoted 10 lines

drivers/cpuidle/governors/menu.c +7 -1

+      * effectively always cause the prediction results to be discarded and
+      * deep idle states to be selected all the time.  That might introduce
+      * unwanted latency into the workload and cause more energy than
+      * necessary to be consumed if the discarded prediction results are
+      * actually accurate, so skip nohz_full CPUs here.
       */
-     if (tick_nohz_tick_stopped() && predicted_ns < TICK_NSEC)
+     if (tick_nohz_tick_stopped() && !tick_nohz_full_cpu(dev->cpu) &&
+         predicted_ns < TICK_NSEC)
              predicted_ns = data->next_timer_ns;

So, when !tick_nohz_full_cpu(dev->cpu), what is the purpose of this tick stopped special case?

Is it because the next dynamic tick is a better prediction than the typical interval once the tick is stopped?

Does that mean we might become more "pessimistic" concerning the predicted idle time for nohz_full CPUs?

I guess too shallow C-states are still better than too deep but there should be a word about that introduced side effect (if any).

Quoted 5 lines · Frederic Weisbecker

So, when !tick_nohz_full_cpu(dev->cpu), what is the purpose of this tick stopped special case?

Is it because the next dynamic tick is a better prediction than the typical interval once the tick is stopped?

When !tick_nohz_full_cpu(dev->cpu), the tick is a safety net against getting stuck in a shallow idle state for too long. In that case, if the tick is stopped, the safety net is not there and it is better to use a deep state.

However, data->next_timer_ns is a lower limit for the idle state target residency because this is when the next timer is going to trigger.

Quoted 3 lines

Does that mean we might become more "pessimistic" concerning the predicted idle time for nohz_full CPUs?

Yes, and not just we might, but we do unless the idle periods in the workload are "long".

Quoted 3 lines

I guess too shallow C-states are still better than too deep but there should be a word about that introduced side effect (if any).

Yeah, I agree.

That said, on a nohz_full CPU there is no safety net against getting stuck in a shallow idle state because the tick is not present. That's why currently the governors don't allow shallow states to be used on nohz_full CPUs.

The lack of a safety net is generally not a problem when the CPU has been isolated to run something doing real work all the time, with possible idle periods in the workload, but there are people who isolate CPUs for energy-saving reasons and don't run anything on them on purpose. For those folks, the current behavior to select deep idle states every time is actually desirable.

So there are two use cases that cannot be addressed at once and I'm thinking about adding a control knob to allow the user to decide which way to go.

Quoted 11 lines · Rafael J. Wysocki

Quoted 5 lines · Frederic Weisbecker

So, when !tick_nohz_full_cpu(dev->cpu), what is the purpose of this tick stopped special case?

Is it because the next dynamic tick is a better prediction than the typical interval once the tick is stopped?

When !tick_nohz_full_cpu(dev->cpu), the tick is a safety net against getting stuck in a shallow idle state for too long. In that case, if the tick is stopped, the safety net is not there and it is better to use a deep state.

Right.

Quoted 3 lines

However, data->next_timer_ns is a lower limit for the idle state target residency because this is when the next timer is going to trigger.

Ok.

Quoted 6 lines

Quoted 3 lines

Does that mean we might become more "pessimistic" concerning the predicted idle time for nohz_full CPUs?

Yes, and not just we might, but we do unless the idle periods in the workload are "long".

Ok.

Quoted 11 lines

Quoted 3 lines

I guess too shallow C-states are still better than too deep but there should be a word about that introduced side effect (if any).

Yeah, I agree.

That said, on a nohz_full CPU there is no safety net against getting stuck in a shallow idle state because the tick is not present. That's why currently the governors don't allow shallow states to be used on nohz_full CPUs.

Quoted 6 lines

The lack of a safety net is generally not a problem when the CPU has been isolated to run something doing real work all the time, with possible idle periods in the workload, but there are people who isolate CPUs for energy-saving reasons and don't run anything on them on purpose. For those folks, the current behavior to select deep idle states every time is actually desirable.

So far I haven't heard from anybody using nohz_full for powersavings. If you have I'd be curious about it.

Whether a task runs tickless or not, it
still runs and the CPU isn't sleeping. Also CPU 0 stays periodic on nohz_full, which alone is a problem for powersaving but also prevents a whole package from entering low power mode on NUMA.

Let's say it not optimized toward powersaving...

Quoted 3 lines

So there are two use cases that cannot be addressed at once and I'm thinking about adding a control knob to allow the user to decide which way to go.

I'm tempted to say we should focus on having not too deep states, at the expense of having too shallow. Of course I'm not entirely comfortable with the idea because nohz_full CPUs may be idle for a while on some workloads. And everyone deserves a rest at some point after a long day.

I guess force restarting the tick upon idle entry would probably be bad for tiny idle round-trips?

As for such a knob, I'm not sure anybody would use it.

Hi Rafael,

I have a question regarding this patch: is it planned for upstream integration, or is there a newer/improved version in the works?

I'm asking because I've been working on a related optimization in the same code path. My patch [1] takes a different approach to the same problem area -- instead of skipping the override entirely for nohz_full CPUs, it changes:

    predicted_ns = data->next_timer_ns;

to:

    predicted_ns = min(predicted_ns, data->next_timer_ns);

The idea is to prevent selecting excessively deep C-states when the prediction is short but the next timer is distant, which on platforms like Sapphire Rapids with high exit latencies (150us+) can cause significant latency spikes.

I notice that our patches touch the same block but address slightly different aspects:

- Your patch: prevents the override from firing on nohz_full CPUs where tick_stopped is always true, avoiding systematically discarding the prediction.
- My patch: when the override does fire, uses min() instead of unconditional replacement to preserve information from the prediction.

These two fixes could potentially be complementary. However, my patch is still under investigation due to limited hardware availability for collecting more data across different platforms.

I'd appreciate your thoughts on whether these approaches could be combined, or if your patch already addresses the use cases I'm seeing.

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

This change follows an analogous modification of the menu governor [1].

Namely, when the governor runs on a nohz_full CPU and there are no user space timers in the workload on that CPU, it ends up selecting idle states with target residency values above TICK_NSEC, or the deepest enabled idle state in the absence of any of those, all the time due to a tick_nohz_tick_stopped() check designed for running on CPUs where the tick is not permanently disabled. In that case, the fact that the tick has been stopped means that the CPU was expected to be idle sufficiently long previously, so it is not unreasonable to expect it to be idle sufficiently long again, but this inference does not apply to nohz_full CPUs.

In some cases, latency in the workload grows undesirably as a result of selecting overly deep idle states, and the workload may also consume more energy than necessary if the CPU does not spend enough time in the selected deep idle state.

Address this by amending the tick_nohz_tick_stopped() check in question with a tick_nohz_full_cpu() one to avoid effectively ignoring all shallow idle states on nohz_full CPUs. While doing so introduces a risk of getting stuck in a shallow idle state for a long time, that only affects energy efficiently, but the current behavior potentially hurts both energy efficiency and performance that is arguably the priority for nohz_full CPUs.

While at it, add a comment explaining the logic in teo_state_ok().

Link https://lore.kernel.org/linux-pm/[email protected]/ [1] open here

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

---
 drivers/cpuidle/governors/teo.c |   18 +++++++++++++-----
 1 file changed, 13 insertions(+), 5 deletions(-)

drivers/cpuidle/governors/teo.c +13 -5

@@ -227,9 +227,17 @@
        cpu_data->total += PULSE;
 }
 
-static bool teo_state_ok(int i, struct cpuidle_driver *drv)
+static bool teo_state_ok(int i, struct cpuidle_driver *drv, struct cpuidle_device *dev)
 {
-       return !tick_nohz_tick_stopped() ||
+       /*
+        * If the scheduler tick has been stopped already, avoid selecting idle
+        * states with target residency below the tick period length under the
+        * assumption that the CPU is likely to be idle sufficiently long for
+        * the tick to be stopped again (or the tick would not have been
+        * stopped previously in the first place).  However, do not do that on
+        * nohz_full CPUs where the above assumption does not hold.
+        */
+       return !tick_nohz_tick_stopped() || tick_nohz_full_cpu(dev->cpu) ||
                drv->states[i].target_residency_ns >= TICK_NSEC;
 }
 

@@ -379,7 +387,7 @@
                                 * shallow or disabled, in which case take the
                                 * first enabled state that is deep enough.
                                 */
-                               if (teo_state_ok(i, drv) &&
+                               if (teo_state_ok(i, drv, dev) &&
                                    !dev->states_usage[i].disable) {
                                        idx = i;
                                        break;

@@ -391,7 +399,7 @@
                        if (dev->states_usage[i].disable)
                                continue;
 
-                       if (teo_state_ok(i, drv)) {
+                       if (teo_state_ok(i, drv, dev)) {
                                /*
                                 * The current state is deep enough, but still
                                 * there may be a better one.

@@ -460,7 +468,7 @@
         */
        if (drv->states[idx].target_residency_ns > duration_ns) {
                i = teo_find_shallower_state(drv, dev, idx, duration_ns, false);
-               if (teo_state_ok(i, drv))
+               if (teo_state_ok(i, drv, dev))
                        idx = i;
        }
 

Quoted 21 lines · Rafael J. Wysocki

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

This change follows an analogous modification of the menu governor [1].

Namely, when the governor runs on a nohz_full CPU and there are no user space timers in the workload on that CPU, it ends up selecting idle states with target residency values above TICK_NSEC, or the deepest enabled idle state in the absence of any of those, all the time due to a tick_nohz_tick_stopped() check designed for running on CPUs where the tick is not permanently disabled. In that case, the fact that the tick has been stopped means that the CPU was expected to be idle sufficiently long previously, so it is not unreasonable to expect it to be idle sufficiently long again, but this inference does not apply to nohz_full CPUs.

In some cases, latency in the workload grows undesirably as a result of selecting overly deep idle states, and the workload may also consume more energy than necessary if the CPU does not spend enough time in the selected deep idle state.

Quoted 7 lines

Address this by amending the tick_nohz_tick_stopped() check in question with a tick_nohz_full_cpu() one to avoid effectively ignoring all shallow idle states on nohz_full CPUs. While doing so introduces a risk of getting stuck in a shallow idle state for a long time, that only affects energy efficiently, but the current behavior potentially hurts both energy efficiency and performance that is arguably the priority for nohz_full CPUs.

This change is likely to break the use case in which CPU isolation is used for power management reasons, to prevent CPUs from running any code and so to save energy.

In that case, going into the deepest state every time on nohz_full CPUs is a feature, so it can't be changed unconditionally.

For this reason, I'm not going to apply this patch and I'm going to drop the menu governor one below.

The only way to allow everyone to do what they want/need I can see would be to add a control knob for adjusting the behavior of cpuidle governors regarding the handling of nohz_full CPUs.

Quoted 40 lines · Rafael J. Wysocki

Quoted 21 lines · Rafael J. Wysocki

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

This change follows an analogous modification of the menu governor [1].

Namely, when the governor runs on a nohz_full CPU and there are no user space timers in the workload on that CPU, it ends up selecting idle states with target residency values above TICK_NSEC, or the deepest enabled idle state in the absence of any of those, all the time due to a tick_nohz_tick_stopped() check designed for running on CPUs where the tick is not permanently disabled. In that case, the fact that the tick has been stopped means that the CPU was expected to be idle sufficiently long previously, so it is not unreasonable to expect it to be idle sufficiently long again, but this inference does not apply to nohz_full CPUs.

In some cases, latency in the workload grows undesirably as a result of selecting overly deep idle states, and the workload may also consume more energy than necessary if the CPU does not spend enough time in the selected deep idle state.

Quoted 7 lines

Address this by amending the tick_nohz_tick_stopped() check in question with a tick_nohz_full_cpu() one to avoid effectively ignoring all shallow idle states on nohz_full CPUs. While doing so introduces a risk of getting stuck in a shallow idle state for a long time, that only affects energy efficiently, but the current behavior potentially hurts both energy efficiency and performance that is arguably the priority for nohz_full CPUs.

This change is likely to break the use case in which CPU isolation is used for power management reasons, to prevent CPUs from running any code and so to save energy.

In that case, going into the deepest state every time on nohz_full CPUs is a feature, so it can't be changed unconditionally.

For this reason, I'm not going to apply this patch and I'm going to drop the menu governor one below.

Quoted 3 lines

The only way to allow everyone to do what they want/need I can see would be to add a control knob for adjusting the behavior of cpuidle governors regarding the handling of nohz_full CPUs.

But then what's the advantage instead of just using /sys/devices/system/cpu/cpuX/power/latency
for the nohz_full CPUs (if you don't want the current 'over-eagerly selecting deepest state on nohz_full')?