Re: [RFC PATCH 2/2] CPUfreq ondemand: handle QoS request on DVFS response latency

[Pavel Machek <pavel@ucw.cz>]

On Wed 2012-02-22 17:03:35, MyungJoo Ham wrote:
> With QoS class, DVFS_RESPONSE_LATENCY, users (device drivers and
> userspace processes) may express the desired maximum response latency
> from DVFS mechanisms such as CPUfreq's ondemand governors. Based on such
> QoS requests, the ondemand governor may flexibly adjust sampling rate
> accordingly unless it goes below the min_sampling_rate.
> 
> The benefit of having DVFS_RESPONSE_LATENCY is to have faster response
> from user inputs (mouse clicks, keyboard inputs, touchscreen touches,
> and others) without increasing frequency unconditionally. Because some
> input events may not require any performance increases, increasing the
> frequency unconditionally for inputs may simply consume too much energy.
> Adjusting sampling rate based on user inputs enabled to increase
> frequency with less latency if it requires and not to increase frequency
> if it does not require.
> 
> Signed-off-by: MyungJoo Ham <myungjoo.ham@samsung.com>
> Signed-off-by: Kyungmin Park <kyungmin.park@samsung.com>
> 
> --
> This patch depends on the patch
> "PM / QoS: Introduce new classes: DMA-Throughput and DVFS-Latency".
> and the patch
> "CPUfreq ondemand: update sampling rate without waiting for next
> sampling"
> ---
>  drivers/cpufreq/cpufreq_ondemand.c |  108 ++++++++++++++++++++++++++++++++----
>  1 files changed, 96 insertions(+), 12 deletions(-)
> 
> diff --git a/drivers/cpufreq/cpufreq_ondemand.c b/drivers/cpufreq/cpufreq_ondemand.c
> index 2d66649..b9188f1 100644
> --- a/drivers/cpufreq/cpufreq_ondemand.c
> +++ b/drivers/cpufreq/cpufreq_ondemand.c
> @@ -22,6 +22,7 @@
>  #include <linux/tick.h>
>  #include <linux/ktime.h>
>  #include <linux/sched.h>
> +#include <linux/pm_qos.h>
>  
>  /*
>   * dbs is used in this file as a shortform for demandbased switching
> @@ -93,6 +94,7 @@ struct cpu_dbs_info_s {
>  	 * when user is changing the governor or limits.
>  	 */
>  	struct mutex timer_mutex;
> +	bool activated; /* dbs_timer_init is in effect */
>  };
>  static DEFINE_PER_CPU(struct cpu_dbs_info_s, od_cpu_dbs_info);
>  
> @@ -111,6 +113,8 @@ static struct dbs_tuners {
>  	unsigned int sampling_down_factor;
>  	unsigned int powersave_bias;
>  	unsigned int io_is_busy;
> +	struct notifier_block dvfs_lat_qos_db;
> +	unsigned int dvfs_lat_qos_wants;
>  } dbs_tuners_ins = {
>  	.up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
>  	.sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
> @@ -164,6 +168,23 @@ static inline cputime64_t get_cpu_iowait_time(unsigned int cpu, cputime64_t *wal
>  }
>  
>  /*
> + * Find right sampling rate based on sampling_rate and
> + * QoS requests on dvfs latency.
> + */
> +static unsigned int effective_sampling_rate(void)
> +{
> +	unsigned int effective;
> +
> +	if (dbs_tuners_ins.dvfs_lat_qos_wants)
> +		effective = min(dbs_tuners_ins.dvfs_lat_qos_wants,
> +				dbs_tuners_ins.sampling_rate);
> +	else
> +		effective = dbs_tuners_ins.sampling_rate;
> +
> +	return max(effective, min_sampling_rate);
> +}
> +
> +/*
>   * Find right freq to be set now with powersave_bias on.
>   * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
>   * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
> @@ -207,7 +228,7 @@ static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
>  		dbs_info->freq_lo_jiffies = 0;
>  		return freq_lo;
>  	}
> -	jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
> +	jiffies_total = usecs_to_jiffies(effective_sampling_rate());
>  	jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
>  	jiffies_hi += ((freq_hi - freq_lo) / 2);
>  	jiffies_hi /= (freq_hi - freq_lo);
> @@ -259,7 +280,8 @@ show_one(powersave_bias, powersave_bias);
>  
>  /**
>   * update_sampling_rate - update sampling rate effective immediately if needed.
> - * @new_rate: new sampling rate
> + * @new_rate: new sampling rate. if it is 0, regard sampling rate is not
> + *		changed and assume that qos request value is changed.
>   *
>   * If new rate is smaller than the old, simply updaing
>   * dbs_tuners_int.sampling_rate might not be appropriate. For example,
> @@ -273,9 +295,13 @@ show_one(powersave_bias, powersave_bias);
>  static void update_sampling_rate(unsigned int new_rate)
>  {
>  	int cpu;
> +	unsigned int effective;
> +
> +
> +	if (new_rate)
> +		dbs_tuners_ins.sampling_rate = max(new_rate, min_sampling_rate);
>  
> -	dbs_tuners_ins.sampling_rate = new_rate
> -				     = max(new_rate, min_sampling_rate);
> +	effective = effective_sampling_rate();
>  
>  	for_each_online_cpu(cpu) {
>  		struct cpufreq_policy *policy;
> @@ -283,21 +309,31 @@ static void update_sampling_rate(unsigned int new_rate)
>  		struct timer_list *timer;
>  		unsigned long appointed_at;
>  
> +		/*
> +		 * mutex_destory(&dbs_info->timer_mutex) should not happen
> +		 * in this context.
> +		 */
> +		mutex_lock(&dbs_mutex);
> +
>  		policy = cpufreq_cpu_get(cpu);
>  		if (!policy)
> -			continue;
> +			goto next;
>  		dbs_info = &per_cpu(od_cpu_dbs_info, policy->cpu);
>  		cpufreq_cpu_put(policy);
>  
> +		/* timer_mutex destroyed or will be destoyed soon */
> +		if (!dbs_info->activated)
> +			goto next;
> +
>  		mutex_lock(&dbs_info->timer_mutex);
>  
>  		if (!delayed_work_pending(&dbs_info->work))
> -			goto next;
> +			goto next_timer_mutex;
>  
>  		timer = &dbs_info->work.timer;
>  		appointed_at = timer->expires;
>  
> -		if (time_before(jiffies + usecs_to_jiffies(new_rate),
> +		if (time_before(jiffies + usecs_to_jiffies(effective),
>  				appointed_at)) {
>  
>  			mutex_unlock(&dbs_info->timer_mutex);
> @@ -305,12 +341,15 @@ static void update_sampling_rate(unsigned int new_rate)
>  			mutex_lock(&dbs_info->timer_mutex);
>  
>  			schedule_delayed_work_on(dbs_info->cpu, &dbs_info->work,
> -						 usecs_to_jiffies(new_rate));
> +						 usecs_to_jiffies(effective));
>  
>  		}
> -next:
> +next_timer_mutex:
>  		mutex_unlock(&dbs_info->timer_mutex);
> +next:
> +		mutex_unlock(&dbs_mutex);
>  	}
> +
>  }

I don't think gotos are helpful here. Can you use normal program
structure or move it to subroutine...?
									Pavel
-- 
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html