Re: [PATCH] CPUIdle: Reevaluate C-states under CPU load to favor deeper C-states

[Kevin Hilman <khilman@ti.com>]

Hi Arjan,

Kevin Hilman <khilman@ti.com> writes:

> From: Nicole Chalhoub <n-chalhoub@ti.com>
>
> While there is CPU load, program a C-state specific one-shot timer in
> order to give CPUidle another opportunity to pick a deeper C-state
> instead of spending potentially long idle times in a shallow C-state.

Any comments on this?

This is an implementation of an idea proposed by you on our previous
attempt[1] to do something similar.

Thanks,

Kevin

[1] http://lkml.org/lkml/2011/4/7/155
>
> Long winded version:
> When going idle with a high load average, CPUidle menu governor will
> decide to pick a shallow C-state since one of the guiding principles
> of the menu governor is "The busier the system, the less impact of
> C-states is acceptable" (taken from cpuidle/governors/menu.c.)
> That makes perfect sense.
>
> However, there are missed power-saving opportunities for bursty
> workloads with long idle times (e.g. MP3 playback.)  Given such a
> workload, because of the load average, CPUidle tends to pick a shallow
> C-state.  Because we also go tickless, this shallow C-state is used
> for the duration of the idle period. If the idle period is long, a
> deeper C state would've resulted in better power savings.
> This patch provides an additional opportuntity for CPUidle to pick a
> deeper C-state by programming a timer (with a C-state specific timeout)
> such that the CPUidle governor will have another opportunity to pick a
> deeper C-state.
>
> Adding this timer for C-state reevaluation improved the load estimation
> on our ARM/OMAP4 platform and increased the time spent in deep C-states
> (~50% of idle time in C-states deeper than C1).  A power saving of ~10mA
> at battery level is observed during MP3 playback on OMAP4/Blaze board.
>
> Signed-off-by: Nicole Chalhoub <n-chalhoub@ti.com>
> Signed-off-by: Kevin Hilman <khilman@ti.com>
> ---
>  drivers/cpuidle/cpuidle.c        |   28 +++++++++++++++++++++++++-
>  drivers/cpuidle/governors/menu.c |   39 ++++++++++++++++++++++++++++++++-----
>  include/linux/cpuidle.h          |    4 +++
>  3 files changed, 63 insertions(+), 8 deletions(-)
>
> diff --git a/drivers/cpuidle/cpuidle.c b/drivers/cpuidle/cpuidle.c
> index 1994885..4b1ac0c 100644
> --- a/drivers/cpuidle/cpuidle.c
> +++ b/drivers/cpuidle/cpuidle.c
> @@ -92,13 +92,33 @@ static void cpuidle_idle_call(void)
>  	target_state->time += (unsigned long long)dev->last_residency;
>  	target_state->usage++;
>  
> -	/* give the governor an opportunity to reflect on the outcome */
> -	if (cpuidle_curr_governor->reflect)
> +	hrtimer_cancel(&dev->cstate_timer);
> +
> +	/*
> +	 * Give the governor an opportunity to reflect on the outcome
> +	 * Do not take into account the wakeups due to the hrtimer, they
> +	 * should not impact the predicted idle time.
> +	 */
> +	if ((!dev->hrtimer_expired) && cpuidle_curr_governor->reflect)
>  		cpuidle_curr_governor->reflect(dev);
>  	trace_power_end(0);
>  }
>  
>  /**
> + * cstate_reassessment_timer - interrupt handler of the cstate hrtimer
> + * @handle:	the expired hrtimer
> + */
> +static enum hrtimer_restart cstate_reassessment_timer(struct hrtimer *handle)
> +{
> +	struct cpuidle_device *data =
> +		container_of(handle, struct cpuidle_device, cstate_timer);
> +
> +	data->hrtimer_expired = 1;
> +
> +	return HRTIMER_NORESTART;
> +}
> +
> +/**
>   * cpuidle_install_idle_handler - installs the cpuidle idle loop handler
>   */
>  void cpuidle_install_idle_handler(void)
> @@ -185,6 +205,10 @@ int cpuidle_enable_device(struct cpuidle_device *dev)
>  
>  	dev->enabled = 1;
>  
> +	dev->hrtimer_expired = 0;
> +	hrtimer_init(&dev->cstate_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
> +	dev->cstate_timer.function = cstate_reassessment_timer;
> +
>  	enabled_devices++;
>  	return 0;
>  
> diff --git a/drivers/cpuidle/governors/menu.c b/drivers/cpuidle/governors/menu.c
> index 1b12870..fd54584 100644
> --- a/drivers/cpuidle/governors/menu.c
> +++ b/drivers/cpuidle/governors/menu.c
> @@ -125,10 +125,21 @@ struct menu_device {
>  #define LOAD_INT(x) ((x) >> FSHIFT)
>  #define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100)
>  
> -static int get_loadavg(void)
> +static int get_loadavg(struct cpuidle_device *dev)
>  {
> -	unsigned long this = this_cpu_load();
> +	unsigned long this;
>  
> +	/*
> +	 * this_cpu_load() returns the value of rq->load.weight
> +	 * at the previous scheduler tick and not the current value.
> +	 * If the timer expired, that means we are in idle,there
> +	 * are no more runnable processes in the current queue
> +	 * =>return the current value of rq->load.weight which is 0.
> +	 */
> +	if (dev->hrtimer_expired == 1)
> +		return 0;
> +	else
> +		this = this_cpu_load();
>  
>  	return LOAD_INT(this) * 10 + LOAD_FRAC(this) / 10;
>  }
> @@ -166,13 +177,13 @@ static inline int which_bucket(unsigned int duration)
>   * to be, the higher this multiplier, and thus the higher
>   * the barrier to go to an expensive C state.
>   */
> -static inline int performance_multiplier(void)
> +static inline int performance_multiplier(struct cpuidle_device *dev)
>  {
>  	int mult = 1;
>  
>  	/* for higher loadavg, we are more reluctant */
>  
> -	mult += 2 * get_loadavg();
> +	mult += 2 * get_loadavg(dev);
>  
>  	/* for IO wait tasks (per cpu!) we add 5x each */
>  	mult += 10 * nr_iowait_cpu(smp_processor_id());
> @@ -236,6 +247,7 @@ static int menu_select(struct cpuidle_device *dev)
>  	int latency_req = pm_qos_request(PM_QOS_CPU_DMA_LATENCY);
>  	int i;
>  	int multiplier;
> +	ktime_t timeout;
>  
>  	if (data->needs_update) {
>  		menu_update(dev);
> @@ -256,7 +268,7 @@ static int menu_select(struct cpuidle_device *dev)
>  
>  	data->bucket = which_bucket(data->expected_us);
>  
> -	multiplier = performance_multiplier();
> +	multiplier = performance_multiplier(dev);
>  
>  	/*
>  	 * if the correction factor is 0 (eg first time init or cpu hotplug
> @@ -287,12 +299,27 @@ static int menu_select(struct cpuidle_device *dev)
>  			break;
>  		if (s->exit_latency > latency_req)
>  			break;
> -		if (s->exit_latency * multiplier > data->predicted_us)
> +		if (s->exit_latency * multiplier > data->predicted_us) {
> +			/*
> +			 * Could not enter the next C-state because of a high
> +			 * load. Set a timer in order to check the load again
> +			 * after the timeout expires and re-evaluate cstate.
> +			 */
> +			if (s->hrtimer_timeout != 0 && get_loadavg(dev)) {
> +				timeout =
> +				       ktime_set(0,
> +					   s->hrtimer_timeout * NSEC_PER_USEC);
> +				hrtimer_start(&dev->cstate_timer, timeout,
> +					   HRTIMER_MODE_REL);
> +			}
>  			break;
> +		}
>  		data->exit_us = s->exit_latency;
>  		data->last_state_idx = i;
>  	}
>  
> +	/* Reset hrtimer_expired which is set when the hrtimer fires */
> +	dev->hrtimer_expired = 0;
>  	return data->last_state_idx;
>  }
>  
> diff --git a/include/linux/cpuidle.h b/include/linux/cpuidle.h
> index 55215cc..8d11b52 100644
> --- a/include/linux/cpuidle.h
> +++ b/include/linux/cpuidle.h
> @@ -16,6 +16,7 @@
>  #include <linux/module.h>
>  #include <linux/kobject.h>
>  #include <linux/completion.h>
> +#include <linux/hrtimer.h>
>  
>  #define CPUIDLE_STATE_MAX	8
>  #define CPUIDLE_NAME_LEN	16
> @@ -37,6 +38,7 @@ struct cpuidle_state {
>  	unsigned int	exit_latency; /* in US */
>  	unsigned int	power_usage; /* in mW */
>  	unsigned int	target_residency; /* in US */
> +	unsigned int	hrtimer_timeout; /* in US */
>  
>  	unsigned long long	usage;
>  	unsigned long long	time; /* in US */
> @@ -97,6 +99,8 @@ struct cpuidle_device {
>  	struct completion	kobj_unregister;
>  	void			*governor_data;
>  	struct cpuidle_state	*safe_state;
> +	struct hrtimer          cstate_timer;
> +	unsigned int            hrtimer_expired;
>  };
>  
>  DECLARE_PER_CPU(struct cpuidle_device *, cpuidle_devices);