[PATCH v2 0/1] Forced-wakeup for stop states on Powernv

Currently, the cpuidle governors determine what idle state a idling CPU
should enter into based on heuristics that depend on the idle history on
that CPU. Given that no predictive heuristic is perfect, there are cases
where the governor predicts a shallow idle state, hoping that the CPU will
be busy soon. However, if no new workload is scheduled on that CPU in the
near future, the CPU will end up in the shallow state.

Motivation
----------
In case of POWER, this is problematic, when the predicted state in the
aforementioned scenario is a shallow stop state on a tickless system. As
we might get stuck into shallow states even for hours, in absence of ticks
or interrupts.

To address this, We forcefully wakeup the cpu by setting the decrementer.
The decrementer is set to a value that corresponds with the residency of
the next available state. Thus firing up a timer that will forcefully
wakeup the cpu. Few such iterations will essentially train the governor to
select a deeper state for that cpu, as the timer here corresponds to the
next available cpuidle state residency. Thus, cpu will eventually end up
in the deepest possible state and we won't get stuck in a shallow state
for long duration.

Experiment
----------
For earlier versions when this feature was meat to be only for shallow lite
states, I performed experiments for three scenarios to collect some data.

case 1 :
Without this patch and without tick retained, i.e. in a upstream kernel,
It would spend more than even a second to get out of stop0_lite.

case 2 : With tick retained in a upstream kernel -

Generally, we have a sched tick at 4ms(CONF_HZ = 250). Ideally I expected
it to take 8 sched tick to get out of stop0_lite. Experimentally,
observation was

=========================================================
sample          min            max           99percentile
20              4ms            12ms          4ms
=========================================================

It would take atleast one sched tick to get out of stop0_lite.

case 2 :  With this patch (not stopping tick, but explicitly queuing a
          timer)

============================================================
sample          min             max             99percentile
============================================================
20              144us           192us           144us
============================================================


Description of current implementation
-------------------------------------

We calculate timeout for the current idle state as the residency value
of the next available idle state. If the decrementer is set to be
greater than this timeout, we update the decrementer value with the
residency of next available idle state. Thus, essentially training the
governor to select the next available deeper state until we reach the
deepest state. Hence, we won't get stuck unnecessarily in shallow states
for longer duration.

--------------------------------
v1 of auto-promotion : https://lkml.org/lkml/2019/3/22/58 This patch was
implemented only for shallow lite state in generic cpuidle driver.

v2 of auto-promotion : Removed timeout_needed and rebased to current
upstream kernel

Then, 
v1 of forced-wakeup : Moved the code to cpuidle powernv driver and started
as forced wakeup instead of auto-promotion

v2 of forced-wakeup : Extended the forced wakeup logic for all states.
Setting the decrementer instead of queuing up a hrtimer to implement the
logic.

Abhishek Goel (1):
  cpuidle-powernv : forced wakeup for stop states

 drivers/cpuidle/cpuidle-powernv.c | 38 +++++++++++++++++++++++++++++++
 1 file changed, 38 insertions(+)

-- 
2.17.1

Currently, the cpuidle governors determine what idle state a idling CPU
should enter into based on heuristics that depend on the idle history on
that CPU. Given that no predictive heuristic is perfect, there are cases
where the governor predicts a shallow idle state, hoping that the CPU will
be busy soon. However, if no new workload is scheduled on that CPU in the
near future, the CPU may end up in the shallow state.

This is problematic, when the predicted state in the aforementioned
scenario is a shallow stop state on a tickless system. As we might get
stuck into shallow states for hours, in absence of ticks or interrupts.

To address this, We forcefully wakeup the cpu by setting the
decrementer. The decrementer is set to a value that corresponds with the
residency of the next available state. Thus firing up a timer that will
forcefully wakeup the cpu. Few such iterations will essentially train the
governor to select a deeper state for that cpu, as the timer here
corresponds to the next available cpuidle state residency. Thus, cpu will
eventually end up in the deepest possible state.

Signed-off-by: Abhishek Goel <redacted>
---

Auto-promotion
 v1 : started as auto promotion logic for cpuidle states in generic
driver
 v2 : Removed timeout_needed and rebased the code to upstream kernel
Forced-wakeup
 v1 : New patch with name of forced wakeup started
 v2 : Extending the forced wakeup logic for all states. Setting the
decrementer instead of queuing up a hrtimer to implement the logic.

 drivers/cpuidle/cpuidle-powernv.c | 38 +++++++++++++++++++++++++++++++
 1 file changed, 38 insertions(+)

diff --git a/drivers/cpuidle/cpuidle-powernv.c b/drivers/cpuidle/cpuidle-powernv.c
index 84b1ebe212b3..bc9ca18ae7e3 100644
--- a/drivers/cpuidle/cpuidle-powernv.c
+++ b/drivers/cpuidle/cpuidle-powernv.c

@@ -46,6 +46,26 @@ static struct stop_psscr_table stop_psscr_table[CPUIDLE_STATE_MAX] __read_mostly
 static u64 default_snooze_timeout __read_mostly;
 static bool snooze_timeout_en __read_mostly;
 
+static u64 forced_wakeup_timeout(struct cpuidle_device *dev,
+				 struct cpuidle_driver *drv,
+				 int index)
+{
+	int i;
+
+	for (i = index + 1; i < drv->state_count; i++) {
+		struct cpuidle_state *s = &drv->states[i];
+		struct cpuidle_state_usage *su = &dev->states_usage[i];
+
+		if (s->disabled || su->disable)
+			continue;
+
+		return (s->target_residency + 2 * s->exit_latency) *
+			tb_ticks_per_usec;
+	}
+
+	return 0;
+}
+
 static u64 get_snooze_timeout(struct cpuidle_device *dev,
 			      struct cpuidle_driver *drv,
 			      int index)

@@ -144,8 +164,26 @@ static int stop_loop(struct cpuidle_device *dev,
 		     struct cpuidle_driver *drv,
 		     int index)
 {
+	u64 dec_expiry_tb, dec, timeout_tb, forced_wakeup;
+
+	dec = mfspr(SPRN_DEC);
+	timeout_tb = forced_wakeup_timeout(dev, drv, index);
+	forced_wakeup = 0;
+
+	if (timeout_tb && timeout_tb < dec) {
+		forced_wakeup = 1;
+		dec_expiry_tb = mftb() + dec;
+	}
+
+	if (forced_wakeup)
+		mtspr(SPRN_DEC, timeout_tb);
+
 	power9_idle_type(stop_psscr_table[index].val,
 			 stop_psscr_table[index].mask);
+
+	if (forced_wakeup)
+		mtspr(SPRN_DEC, dec_expiry_tb - mftb());
+
 	return index;
 }
 

-- 
2.17.1

Abhishek Goel's on June 17, 2019 7:56 pm:

It would be nice to not have this kind of loop iteration in the
idle fast path. Can we add a flag or something to the idle state?

The compiler probably can't optimise away the SPR manipulations so try
to avoid them if possible.

+
+	if (forced_wakeup)
+		mtspr(SPRN_DEC, timeout_tb);

This should just be put in the above 'if'.

+
 	power9_idle_type(stop_psscr_table[index].val,
 			 stop_psscr_table[index].mask);
+
+	if (forced_wakeup)
+		mtspr(SPRN_DEC, dec_expiry_tb - mftb());

This will sometimes go negative and result in another timer interrupt.

It also breaks irq work (which can be set here by machine check I
believe.

May need to implement some timer code to do this for you.

static void reset_dec_after_idle(void)
{
	u64 now;
        u64 *next_tb;

	if (test_irq_work_pending())
		return;
	now = mftb;
	next_tb = this_cpu_ptr(&decrementers_next_tb);

	if (now >= *next_tb)
		return;
	set_dec(*next_tb - now);
	if (test_irq_work_pending())
		set_dec(1);
}

Something vaguely like that. See timer_interrupt().

Thanks,
Nick

Hi Nick,

Thanks for the review. Some replies below.

On 06/19/2019 09:53 AM, Nicholas Piggin wrote:

Abhishek Goel's on June 17, 2019 7:56 pm:

quoted

Currently, the cpuidle governors determine what idle state a idling CPU
should enter into based on heuristics that depend on the idle history on
that CPU. Given that no predictive heuristic is perfect, there are cases
where the governor predicts a shallow idle state, hoping that the CPU will
be busy soon. However, if no new workload is scheduled on that CPU in the
near future, the CPU may end up in the shallow state.

This is problematic, when the predicted state in the aforementioned
scenario is a shallow stop state on a tickless system. As we might get
stuck into shallow states for hours, in absence of ticks or interrupts.

To address this, We forcefully wakeup the cpu by setting the
decrementer. The decrementer is set to a value that corresponds with the
residency of the next available state. Thus firing up a timer that will
forcefully wakeup the cpu. Few such iterations will essentially train the
governor to select a deeper state for that cpu, as the timer here
corresponds to the next available cpuidle state residency. Thus, cpu will
eventually end up in the deepest possible state.

Signed-off-by: Abhishek Goel <redacted>
---

Auto-promotion
  v1 : started as auto promotion logic for cpuidle states in generic
driver
  v2 : Removed timeout_needed and rebased the code to upstream kernel
Forced-wakeup
  v1 : New patch with name of forced wakeup started
  v2 : Extending the forced wakeup logic for all states. Setting the
decrementer instead of queuing up a hrtimer to implement the logic.

  drivers/cpuidle/cpuidle-powernv.c | 38 +++++++++++++++++++++++++++++++
  1 file changed, 38 insertions(+)

diff --git a/drivers/cpuidle/cpuidle-powernv.c b/drivers/cpuidle/cpuidle-powernv.c
index 84b1ebe212b3..bc9ca18ae7e3 100644
--- a/drivers/cpuidle/cpuidle-powernv.c
+++ b/drivers/cpuidle/cpuidle-powernv.c

@@ -46,6 +46,26 @@ static struct stop_psscr_table stop_psscr_table[CPUIDLE_STATE_MAX] __read_mostly
  static u64 default_snooze_timeout __read_mostly;
  static bool snooze_timeout_en __read_mostly;
  
+static u64 forced_wakeup_timeout(struct cpuidle_device *dev,
+				 struct cpuidle_driver *drv,
+				 int index)
+{
+	int i;
+
+	for (i = index + 1; i < drv->state_count; i++) {
+		struct cpuidle_state *s = &drv->states[i];
+		struct cpuidle_state_usage *su = &dev->states_usage[i];
+
+		if (s->disabled || su->disable)
+			continue;
+
+		return (s->target_residency + 2 * s->exit_latency) *
+			tb_ticks_per_usec;
+	}
+
+	return 0;
+}

It would be nice to not have this kind of loop iteration in the
idle fast path. Can we add a flag or something to the idle state?

Currently, we do not have any callback notification or some feedback that
notifies the driver everytime some state is enabled/disabled. So we have
to parse everytime to get the next enabled state. Are you suggesting to
add something like next_enabled_state in cpuidle state structure itself
which will be updated when a state is enabled or disabled?

quoted

+
  static u64 get_snooze_timeout(struct cpuidle_device *dev,
  			      struct cpuidle_driver *drv,
  			      int index)

@@ -144,8 +164,26 @@ static int stop_loop(struct cpuidle_device *dev,
  		     struct cpuidle_driver *drv,
  		     int index)
  {
+	u64 dec_expiry_tb, dec, timeout_tb, forced_wakeup;
+
+	dec = mfspr(SPRN_DEC);
+	timeout_tb = forced_wakeup_timeout(dev, drv, index);
+	forced_wakeup = 0;
+
+	if (timeout_tb && timeout_tb < dec) {
+		forced_wakeup = 1;
+		dec_expiry_tb = mftb() + dec;
+	}

The compiler probably can't optimise away the SPR manipulations so try
to avoid them if possible.

Are you suggesting something like set_dec_before_idle?(in line with
what you have suggested to do after idle, reset_dec_after_idle)

quoted

+
+	if (forced_wakeup)
+		mtspr(SPRN_DEC, timeout_tb);

This should just be put in the above 'if'.

Fair point.

quoted

+
  	power9_idle_type(stop_psscr_table[index].val,
  			 stop_psscr_table[index].mask);
+
+	if (forced_wakeup)
+		mtspr(SPRN_DEC, dec_expiry_tb - mftb());

This will sometimes go negative and result in another timer interrupt.

It also breaks irq work (which can be set here by machine check I
believe.

May need to implement some timer code to do this for you.

static void reset_dec_after_idle(void)
{
	u64 now;
         u64 *next_tb;

	if (test_irq_work_pending())
		return;
	now = mftb;
	next_tb = this_cpu_ptr(&decrementers_next_tb);

	if (now >= *next_tb)
		return;
	set_dec(*next_tb - now);
	if (test_irq_work_pending())
		set_dec(1);
}

Something vaguely like that. See timer_interrupt().

Ah, Okay. Will go through timer_interrupt().

Thanks,
Nick

Thanks,
Abhishek

Abhishek's on June 19, 2019 7:08 pm:

Hi Nick,

Thanks for the review. Some replies below.

On 06/19/2019 09:53 AM, Nicholas Piggin wrote:

quoted

Abhishek Goel's on June 17, 2019 7:56 pm:

quoted

Currently, the cpuidle governors determine what idle state a idling CPU
should enter into based on heuristics that depend on the idle history on
that CPU. Given that no predictive heuristic is perfect, there are cases
where the governor predicts a shallow idle state, hoping that the CPU will
be busy soon. However, if no new workload is scheduled on that CPU in the
near future, the CPU may end up in the shallow state.

This is problematic, when the predicted state in the aforementioned
scenario is a shallow stop state on a tickless system. As we might get
stuck into shallow states for hours, in absence of ticks or interrupts.

To address this, We forcefully wakeup the cpu by setting the
decrementer. The decrementer is set to a value that corresponds with the
residency of the next available state. Thus firing up a timer that will
forcefully wakeup the cpu. Few such iterations will essentially train the
governor to select a deeper state for that cpu, as the timer here
corresponds to the next available cpuidle state residency. Thus, cpu will
eventually end up in the deepest possible state.

Signed-off-by: Abhishek Goel <redacted>
---

Auto-promotion
  v1 : started as auto promotion logic for cpuidle states in generic
driver
  v2 : Removed timeout_needed and rebased the code to upstream kernel
Forced-wakeup
  v1 : New patch with name of forced wakeup started
  v2 : Extending the forced wakeup logic for all states. Setting the
decrementer instead of queuing up a hrtimer to implement the logic.

  drivers/cpuidle/cpuidle-powernv.c | 38 +++++++++++++++++++++++++++++++
  1 file changed, 38 insertions(+)

diff --git a/drivers/cpuidle/cpuidle-powernv.c b/drivers/cpuidle/cpuidle-powernv.c
index 84b1ebe212b3..bc9ca18ae7e3 100644
--- a/drivers/cpuidle/cpuidle-powernv.c
+++ b/drivers/cpuidle/cpuidle-powernv.c

@@ -46,6 +46,26 @@ static struct stop_psscr_table stop_psscr_table[CPUIDLE_STATE_MAX] __read_mostly
  static u64 default_snooze_timeout __read_mostly;
  static bool snooze_timeout_en __read_mostly;
  
+static u64 forced_wakeup_timeout(struct cpuidle_device *dev,
+				 struct cpuidle_driver *drv,
+				 int index)
+{
+	int i;
+
+	for (i = index + 1; i < drv->state_count; i++) {
+		struct cpuidle_state *s = &drv->states[i];
+		struct cpuidle_state_usage *su = &dev->states_usage[i];
+
+		if (s->disabled || su->disable)
+			continue;
+
+		return (s->target_residency + 2 * s->exit_latency) *
+			tb_ticks_per_usec;
+	}
+
+	return 0;
+}

It would be nice to not have this kind of loop iteration in the
idle fast path. Can we add a flag or something to the idle state?

Currently, we do not have any callback notification or some feedback that
notifies the driver everytime some state is enabled/disabled. So we have
to parse everytime to get the next enabled state.

Ahh, that's why you're doing that.

Are you suggesting to
add something like next_enabled_state in cpuidle state structure itself
which will be updated when a state is enabled or disabled?

Hmm, I guess it normally should not iterate over more than one state
unless some idle states are disabled.

What would have been nice is each state just have its own timeout
field with ticks already calculated, if that could be updated when
a state is enabled or disabled. How hard is that to add to the
cpuidle core?

quoted

quoted

+
  static u64 get_snooze_timeout(struct cpuidle_device *dev,
  			      struct cpuidle_driver *drv,
  			      int index)

@@ -144,8 +164,26 @@ static int stop_loop(struct cpuidle_device *dev,
  		     struct cpuidle_driver *drv,
  		     int index)
  {
+	u64 dec_expiry_tb, dec, timeout_tb, forced_wakeup;
+
+	dec = mfspr(SPRN_DEC);
+	timeout_tb = forced_wakeup_timeout(dev, drv, index);
+	forced_wakeup = 0;
+
+	if (timeout_tb && timeout_tb < dec) {
+		forced_wakeup = 1;
+		dec_expiry_tb = mftb() + dec;
+	}

The compiler probably can't optimise away the SPR manipulations so try
to avoid them if possible.

Are you suggesting something like set_dec_before_idle?(in line with
what you have suggested to do after idle, reset_dec_after_idle)

I should have been clear, I meant don't mfspr(SPRN_DEC) until you
have tested timeout_tb.

quoted

quoted

+
+	if (forced_wakeup)
+		mtspr(SPRN_DEC, timeout_tb);

This should just be put in the above 'if'.

Fair point.

quoted

quoted

+
  	power9_idle_type(stop_psscr_table[index].val,
  			 stop_psscr_table[index].mask);
+
+	if (forced_wakeup)
+		mtspr(SPRN_DEC, dec_expiry_tb - mftb());

This will sometimes go negative and result in another timer interrupt.

It also breaks irq work (which can be set here by machine check I
believe.

May need to implement some timer code to do this for you.

static void reset_dec_after_idle(void)
{
	u64 now;
         u64 *next_tb;

	if (test_irq_work_pending())
		return;
	now = mftb;
	next_tb = this_cpu_ptr(&decrementers_next_tb);

	if (now >= *next_tb)
		return;
	set_dec(*next_tb - now);
	if (test_irq_work_pending())
		set_dec(1);
}

Something vaguely like that. See timer_interrupt().

Ah, Okay. Will go through timer_interrupt().

Thanks,
Nick

Hi Nick,


On 06/19/2019 03:39 PM, Nicholas Piggin wrote:

Abhishek's on June 19, 2019 7:08 pm:

quoted

Hi Nick,

Thanks for the review. Some replies below.

On 06/19/2019 09:53 AM, Nicholas Piggin wrote:

quoted

Abhishek Goel's on June 17, 2019 7:56 pm:

quoted

Currently, the cpuidle governors determine what idle state a idling CPU
should enter into based on heuristics that depend on the idle history on
that CPU. Given that no predictive heuristic is perfect, there are cases
where the governor predicts a shallow idle state, hoping that the CPU will
be busy soon. However, if no new workload is scheduled on that CPU in the
near future, the CPU may end up in the shallow state.

This is problematic, when the predicted state in the aforementioned
scenario is a shallow stop state on a tickless system. As we might get
stuck into shallow states for hours, in absence of ticks or interrupts.

To address this, We forcefully wakeup the cpu by setting the
decrementer. The decrementer is set to a value that corresponds with the
residency of the next available state. Thus firing up a timer that will
forcefully wakeup the cpu. Few such iterations will essentially train the
governor to select a deeper state for that cpu, as the timer here
corresponds to the next available cpuidle state residency. Thus, cpu will
eventually end up in the deepest possible state.

Signed-off-by: Abhishek Goel <redacted>
---

Auto-promotion
   v1 : started as auto promotion logic for cpuidle states in generic
driver
   v2 : Removed timeout_needed and rebased the code to upstream kernel
Forced-wakeup
   v1 : New patch with name of forced wakeup started
   v2 : Extending the forced wakeup logic for all states. Setting the
decrementer instead of queuing up a hrtimer to implement the logic.

   drivers/cpuidle/cpuidle-powernv.c | 38 +++++++++++++++++++++++++++++++
   1 file changed, 38 insertions(+)

diff --git a/drivers/cpuidle/cpuidle-powernv.c b/drivers/cpuidle/cpuidle-powernv.c
index 84b1ebe212b3..bc9ca18ae7e3 100644
--- a/drivers/cpuidle/cpuidle-powernv.c
+++ b/drivers/cpuidle/cpuidle-powernv.c

@@ -46,6 +46,26 @@ static struct stop_psscr_table stop_psscr_table[CPUIDLE_STATE_MAX] __read_mostly
   static u64 default_snooze_timeout __read_mostly;
   static bool snooze_timeout_en __read_mostly;
   
+static u64 forced_wakeup_timeout(struct cpuidle_device *dev,
+				 struct cpuidle_driver *drv,
+				 int index)
+{
+	int i;
+
+	for (i = index + 1; i < drv->state_count; i++) {
+		struct cpuidle_state *s = &drv->states[i];
+		struct cpuidle_state_usage *su = &dev->states_usage[i];
+
+		if (s->disabled || su->disable)
+			continue;
+
+		return (s->target_residency + 2 * s->exit_latency) *
+			tb_ticks_per_usec;
+	}
+
+	return 0;
+}

It would be nice to not have this kind of loop iteration in the
idle fast path. Can we add a flag or something to the idle state?

Currently, we do not have any callback notification or some feedback that
notifies the driver everytime some state is enabled/disabled. So we have
to parse everytime to get the next enabled state.

Ahh, that's why you're doing that.

quoted

Are you suggesting to
add something like next_enabled_state in cpuidle state structure itself
which will be updated when a state is enabled or disabled?

Hmm, I guess it normally should not iterate over more than one state
unless some idle states are disabled.

What would have been nice is each state just have its own timeout
field with ticks already calculated, if that could be updated when
a state is enabled or disabled. How hard is that to add to the
cpuidle core?

I have implemented a prototype which does what you have asked for. Added
a  disable_callback which will update timeout whenever a state is 
enabled or
disabled. But It would mean adding some code to cpuidle.h and 
cpuidle/sysfs.c.
If that is not an issue, should I go ahead and post it?

quoted

quoted

quoted

+
   static u64 get_snooze_timeout(struct cpuidle_device *dev,
   			      struct cpuidle_driver *drv,
   			      int index)

@@ -144,8 +164,26 @@ static int stop_loop(struct cpuidle_device *dev,
   		     struct cpuidle_driver *drv,
   		     int index)
   {
+	u64 dec_expiry_tb, dec, timeout_tb, forced_wakeup;
+
+	dec = mfspr(SPRN_DEC);
+	timeout_tb = forced_wakeup_timeout(dev, drv, index);
+	forced_wakeup = 0;
+
+	if (timeout_tb && timeout_tb < dec) {
+		forced_wakeup = 1;
+		dec_expiry_tb = mftb() + dec;
+	}

The compiler probably can't optimise away the SPR manipulations so try
to avoid them if possible.

Are you suggesting something like set_dec_before_idle?(in line with
what you have suggested to do after idle, reset_dec_after_idle)

I should have been clear, I meant don't mfspr(SPRN_DEC) until you
have tested timeout_tb.

quoted

quoted

quoted

+
+	if (forced_wakeup)
+		mtspr(SPRN_DEC, timeout_tb);

This should just be put in the above 'if'.

Fair point.

quoted

quoted

+
   	power9_idle_type(stop_psscr_table[index].val,
   			 stop_psscr_table[index].mask);
+
+	if (forced_wakeup)
+		mtspr(SPRN_DEC, dec_expiry_tb - mftb());

This will sometimes go negative and result in another timer interrupt.

It also breaks irq work (which can be set here by machine check I
believe.

May need to implement some timer code to do this for you.

static void reset_dec_after_idle(void)
{
	u64 now;
          u64 *next_tb;

	if (test_irq_work_pending())
		return;
	now = mftb;
	next_tb = this_cpu_ptr(&decrementers_next_tb);

	if (now >= *next_tb)
		return;
	set_dec(*next_tb - now);
	if (test_irq_work_pending())
		set_dec(1);
}

Something vaguely like that. See timer_interrupt().

Ah, Okay. Will go through timer_interrupt().

Thanks,
Nick

Thanks,
Abhishek