linux-kernel-cpufreq_ondemand

/*
 *  drivers/cpufreq/cpufreq_ondemand.c
 *
 *  Copyright (C)  2001 Russell King
 *            (C)  2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
 *                      Jun Nakajima <jun.nakajima@intel.com>
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License version 2 as
 * published by the Free Software Foundation.
 */

#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt

#include <linux/cpufreq.h>
#include <linux/init.h>
#include <linux/kernel.h>
#include <linux/kernel_stat.h>
#include <linux/kobject.h>
#include <linux/module.h>
#include <linux/mutex.h>
#include <linux/percpu-defs.h>
#include <linux/slab.h>
#include <linux/sysfs.h>
#include <linux/tick.h>
#include <linux/types.h>
#include <linux/cpu.h>

#include "cpufreq_governor.h"

/* On-demand governor macros */
#define DEF_FREQUENCY_DOWN_DIFFERENTIAL  (10)
#define DEF_FREQUENCY_UP_THRESHOLD  (80)
#define DEF_SAMPLING_DOWN_FACTOR  (1)
#define MAX_SAMPLING_DOWN_FACTOR  (100000)
#define MICRO_FREQUENCY_DOWN_DIFFERENTIAL (3)
#define MICRO_FREQUENCY_UP_THRESHOLD  (99)
#define MICRO_FREQUENCY_MIN_SAMPLE_RATE  (10000)
#define MIN_FREQUENCY_UP_THRESHOLD  (11)
#define MAX_FREQUENCY_UP_THRESHOLD  (100)

static DEFINE_PER_CPU(struct od_cpu_dbs_info_s, od_cpu_dbs_info);

static struct od_ops od_ops;

#ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
static struct cpufreq_governor cpufreq_gov_ondemand;
#endif

static unsigned int default_powersave_bias;

static void ondemand_powersave_bias_init_cpu(int cpu)
{
 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);

 dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
 dbs_info->freq_lo = 0;
}

/*
 * Not all CPUs want IO time to be accounted as busy; this depends on how
 * efficient idling at a higher frequency/voltage is.
 * Pavel Machek says this is not so for various generations of AMD and old
 * Intel systems.
 * Mike Chan (android.com) claims this is also not true for ARM.
 * Because of this, whitelist specific known (series) of CPUs by default, and
 * leave all others up to the user.
 */
static int should_io_be_busy(void)
{
#if defined(CONFIG_X86)
 /*
  * For Intel, Core 2 (model 15) and later have an efficient idle.
  */
 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
   boot_cpu_data.x86 == 6 &&
   boot_cpu_data.x86_model >= 15)
  return 1;
#endif
 return 0;
}

/*
 * 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.
 */
static unsigned int generic_powersave_bias_target(struct cpufreq_policy *policy,
  unsigned int freq_next, unsigned int relation)
{
 unsigned int freq_req, freq_reduc, freq_avg;
 unsigned int freq_hi, freq_lo;
 unsigned int index = 0;
 unsigned int jiffies_total, jiffies_hi, jiffies_lo;
 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
         policy->cpu);
 struct dbs_data *dbs_data = policy->governor_data;
 struct od_dbs_tuners *od_tuners = dbs_data->tuners;

 if (!dbs_info->freq_table) {
  dbs_info->freq_lo = 0;
  dbs_info->freq_lo_jiffies = 0;
  return freq_next;
 }

 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
   relation, &index);
 freq_req = dbs_info->freq_table[index].frequency;
 freq_reduc = freq_req * od_tuners->powersave_bias / 1000;
 freq_avg = freq_req - freq_reduc;

 /* Find freq bounds for freq_avg in freq_table */
 index = 0;
 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
   CPUFREQ_RELATION_H, &index);
 freq_lo = dbs_info->freq_table[index].frequency;
 index = 0;
 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
   CPUFREQ_RELATION_L, &index);
 freq_hi = dbs_info->freq_table[index].frequency;

 /* Find out how long we have to be in hi and lo freqs */
 if (freq_hi == freq_lo) {
  dbs_info->freq_lo = 0;
  dbs_info->freq_lo_jiffies = 0;
  return freq_lo;
 }
 jiffies_total = usecs_to_jiffies(od_tuners->sampling_rate);
 jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
 jiffies_hi += ((freq_hi - freq_lo) / 2);
 jiffies_hi /= (freq_hi - freq_lo);
 jiffies_lo = jiffies_total - jiffies_hi;
 dbs_info->freq_lo = freq_lo;
 dbs_info->freq_lo_jiffies = jiffies_lo;
 dbs_info->freq_hi_jiffies = jiffies_hi;
 return freq_hi;
}

static void ondemand_powersave_bias_init(void)
{
 int i;
 for_each_online_cpu(i) {
  ondemand_powersave_bias_init_cpu(i);
 }
}

static void dbs_freq_increase(struct cpufreq_policy *p, unsigned int freq)
{
 struct dbs_data *dbs_data = p->governor_data;
 struct od_dbs_tuners *od_tuners = dbs_data->tuners;

 if (od_tuners->powersave_bias)
  freq = od_ops.powersave_bias_target(p, freq,
    CPUFREQ_RELATION_H);
 else if (p->cur == p->max)
  return;

 __cpufreq_driver_target(p, freq, od_tuners->powersave_bias ?
   CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
}

/*
 * Every sampling_rate, we check, if current idle time is less than 20%
 * (default), then we try to increase frequency. Every sampling_rate, we look
 * for the lowest frequency which can sustain the load while keeping idle time
 * over 30%. If such a frequency exist, we try to decrease to this frequency.
 *
 * Any frequency increase takes it to the maximum frequency. Frequency reduction
 * happens at minimum steps of 5% (default) of current frequency
 */
static void od_check_cpu(int cpu, unsigned int load_freq)
{
 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 struct cpufreq_policy *policy = dbs_info->cdbs.cur_policy;
 struct dbs_data *dbs_data = policy->governor_data;
 struct od_dbs_tuners *od_tuners = dbs_data->tuners;

 dbs_info->freq_lo = 0;

 /* Check for frequency increase */
 if (load_freq > od_tuners->up_threshold * policy->cur) {
  /* If switching to max speed, apply sampling_down_factor */
  if (policy->cur < policy->max)
   dbs_info->rate_mult =
    od_tuners->sampling_down_factor;
  dbs_freq_increase(policy, policy->max);
  return;
 }

 /* Check for frequency decrease */
 /* if we cannot reduce the frequency anymore, break out early */
 if (policy->cur == policy->min)
  return;

 /*
  * The optimal frequency is the frequency that is the lowest that can
  * support the current CPU usage without triggering the up policy. To be
  * safe, we focus 10 points under the threshold.
  */
 if (load_freq < od_tuners->adj_up_threshold
   * policy->cur) {
  unsigned int freq_next;
  freq_next = load_freq / od_tuners->adj_up_threshold;

  /* No longer fully busy, reset rate_mult */
  dbs_info->rate_mult = 1;

  if (freq_next < policy->min)
   freq_next = policy->min;

  if (!od_tuners->powersave_bias) {
   __cpufreq_driver_target(policy, freq_next,
     CPUFREQ_RELATION_L);
   return;
  }

  freq_next = od_ops.powersave_bias_target(policy, freq_next,
     CPUFREQ_RELATION_L);
  __cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_L);
 }
}

static void od_dbs_timer(struct work_struct *work)
{
 struct od_cpu_dbs_info_s *dbs_info =
  container_of(work, struct od_cpu_dbs_info_s, cdbs.work.work);
 unsigned int cpu = dbs_info->cdbs.cur_policy->cpu;
 struct od_cpu_dbs_info_s *core_dbs_info = &per_cpu(od_cpu_dbs_info,
   cpu);
 struct dbs_data *dbs_data = dbs_info->cdbs.cur_policy->governor_data;
 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 int delay = 0, sample_type = core_dbs_info->sample_type;
 bool modify_all = true;

 mutex_lock(&core_dbs_info->cdbs.timer_mutex);
 if (!need_load_eval(&core_dbs_info->cdbs, od_tuners->sampling_rate)) {
  modify_all = false;
  goto max_delay;
 }

 /* Common NORMAL_SAMPLE setup */
 core_dbs_info->sample_type = OD_NORMAL_SAMPLE;
 if (sample_type == OD_SUB_SAMPLE) {
  delay = core_dbs_info->freq_lo_jiffies;
  __cpufreq_driver_target(core_dbs_info->cdbs.cur_policy,
    core_dbs_info->freq_lo, CPUFREQ_RELATION_H);
 } else {
  dbs_check_cpu(dbs_data, cpu);
  if (core_dbs_info->freq_lo) {
   /* Setup timer for SUB_SAMPLE */
   core_dbs_info->sample_type = OD_SUB_SAMPLE;
   delay = core_dbs_info->freq_hi_jiffies;
  }
 }

max_delay:
 if (!delay)
  delay = delay_for_sampling_rate(od_tuners->sampling_rate
    * core_dbs_info->rate_mult);

 gov_queue_work(dbs_data, dbs_info->cdbs.cur_policy, delay, modify_all);
 mutex_unlock(&core_dbs_info->cdbs.timer_mutex);
}

/************************** sysfs interface ************************/
static struct common_dbs_data od_dbs_cdata;

/**
 * update_sampling_rate - update sampling rate effective immediately if needed.
 * @new_rate: new sampling rate
 *
 * If new rate is smaller than the old, simply updating
 * dbs_tuners_int.sampling_rate might not be appropriate. For example, if the
 * original sampling_rate was 1 second and the requested new sampling rate is 10
 * ms because the user needs immediate reaction from ondemand governor, but not
 * sure if higher frequency will be required or not, then, the governor may
 * change the sampling rate too late; up to 1 second later. Thus, if we are
 * reducing the sampling rate, we need to make the new value effective
 * immediately.
 */
static void update_sampling_rate(struct dbs_data *dbs_data,
  unsigned int new_rate)
{
 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 int cpu;

 od_tuners->sampling_rate = new_rate = max(new_rate,
   dbs_data->min_sampling_rate);

 for_each_online_cpu(cpu) {
  struct cpufreq_policy *policy;
  struct od_cpu_dbs_info_s *dbs_info;
  unsigned long next_sampling, appointed_at;

  policy = cpufreq_cpu_get(cpu);
  if (!policy)
   continue;
  if (policy->governor != &cpufreq_gov_ondemand) {
   cpufreq_cpu_put(policy);
   continue;
  }
  dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
  cpufreq_cpu_put(policy);

  mutex_lock(&dbs_info->cdbs.timer_mutex);

  if (!delayed_work_pending(&dbs_info->cdbs.work)) {
   mutex_unlock(&dbs_info->cdbs.timer_mutex);
   continue;
  }

  next_sampling = jiffies + usecs_to_jiffies(new_rate);
  appointed_at = dbs_info->cdbs.work.timer.expires;

  if (time_before(next_sampling, appointed_at)) {

   mutex_unlock(&dbs_info->cdbs.timer_mutex);
   cancel_delayed_work_sync(&dbs_info->cdbs.work);
   mutex_lock(&dbs_info->cdbs.timer_mutex);

   gov_queue_work(dbs_data, dbs_info->cdbs.cur_policy,
     usecs_to_jiffies(new_rate), true);

  }
  mutex_unlock(&dbs_info->cdbs.timer_mutex);
 }
}

static ssize_t store_sampling_rate(struct dbs_data *dbs_data, const char *buf,
  size_t count)
{
 unsigned int input;
 int ret;
 ret = sscanf(buf, "%u", &input);
 if (ret != 1)
  return -EINVAL;

 update_sampling_rate(dbs_data, input);
 return count;
}

static ssize_t store_io_is_busy(struct dbs_data *dbs_data, const char *buf,
  size_t count)
{
 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 unsigned int input;
 int ret;
 unsigned int j;

 ret = sscanf(buf, "%u", &input);
 if (ret != 1)
  return -EINVAL;
 od_tuners->io_is_busy = !!input;

 /* we need to re-evaluate prev_cpu_idle */
 for_each_online_cpu(j) {
  struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
         j);
  dbs_info->cdbs.prev_cpu_idle = get_cpu_idle_time(j,
   &dbs_info->cdbs.prev_cpu_wall, od_tuners->io_is_busy);
 }
 return count;
}

static ssize_t store_up_threshold(struct dbs_data *dbs_data, const char *buf,
  size_t count)
{
 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 unsigned int input;
 int ret;
 ret = sscanf(buf, "%u", &input);

 if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
   input < MIN_FREQUENCY_UP_THRESHOLD) {
  return -EINVAL;
 }
 /* Calculate the new adj_up_threshold */
 od_tuners->adj_up_threshold += input;
 od_tuners->adj_up_threshold -= od_tuners->up_threshold;

 od_tuners->up_threshold = input;
 return count;
}

static ssize_t store_sampling_down_factor(struct dbs_data *dbs_data,
  const char *buf, size_t count)
{
 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 unsigned int input, j;
 int ret;
 ret = sscanf(buf, "%u", &input);

 if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
  return -EINVAL;
 od_tuners->sampling_down_factor = input;

 /* Reset down sampling multiplier in case it was active */
 for_each_online_cpu(j) {
  struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
    j);
  dbs_info->rate_mult = 1;
 }
 return count;
}

static ssize_t store_ignore_nice_load(struct dbs_data *dbs_data,
  const char *buf, size_t count)
{
 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 unsigned int input;
 int ret;

 unsigned int j;

 ret = sscanf(buf, "%u", &input);
 if (ret != 1)
  return -EINVAL;

 if (input > 1)
  input = 1;

 if (input == od_tuners->ignore_nice_load) { /* nothing to do */
  return count;
 }
 od_tuners->ignore_nice_load = input;

 /* we need to re-evaluate prev_cpu_idle */
 for_each_online_cpu(j) {
  struct od_cpu_dbs_info_s *dbs_info;
  dbs_info = &per_cpu(od_cpu_dbs_info, j);
  dbs_info->cdbs.prev_cpu_idle = get_cpu_idle_time(j,
   &dbs_info->cdbs.prev_cpu_wall, od_tuners->io_is_busy);
  if (od_tuners->ignore_nice_load)
   dbs_info->cdbs.prev_cpu_nice =
    kcpustat_cpu(j).cpustat[CPUTIME_NICE];

 }
 return count;
}

static ssize_t store_powersave_bias(struct dbs_data *dbs_data, const char *buf,
  size_t count)
{
 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 unsigned int input;
 int ret;
 ret = sscanf(buf, "%u", &input);

 if (ret != 1)
  return -EINVAL;

 if (input > 1000)
  input = 1000;

 od_tuners->powersave_bias = input;
 ondemand_powersave_bias_init();
 return count;
}

show_store_one(od, sampling_rate);
show_store_one(od, io_is_busy);
show_store_one(od, up_threshold);
show_store_one(od, sampling_down_factor);
show_store_one(od, ignore_nice_load);
show_store_one(od, powersave_bias);
declare_show_sampling_rate_min(od);

gov_sys_pol_attr_rw(sampling_rate);
gov_sys_pol_attr_rw(io_is_busy);
gov_sys_pol_attr_rw(up_threshold);
gov_sys_pol_attr_rw(sampling_down_factor);
gov_sys_pol_attr_rw(ignore_nice_load);
gov_sys_pol_attr_rw(powersave_bias);
gov_sys_pol_attr_ro(sampling_rate_min);

static struct attribute *dbs_attributes_gov_sys[] = {
 &sampling_rate_min_gov_sys.attr,
 &sampling_rate_gov_sys.attr,
 &up_threshold_gov_sys.attr,
 &sampling_down_factor_gov_sys.attr,
 &ignore_nice_load_gov_sys.attr,
 &powersave_bias_gov_sys.attr,
 &io_is_busy_gov_sys.attr,
 NULL
};

static struct attribute_group od_attr_group_gov_sys = {
 .attrs = dbs_attributes_gov_sys,
 .name = "ondemand",
};

static struct attribute *dbs_attributes_gov_pol[] = {
 &sampling_rate_min_gov_pol.attr,
 &sampling_rate_gov_pol.attr,
 &up_threshold_gov_pol.attr,
 &sampling_down_factor_gov_pol.attr,
 &ignore_nice_load_gov_pol.attr,
 &powersave_bias_gov_pol.attr,
 &io_is_busy_gov_pol.attr,
 NULL
};

static struct attribute_group od_attr_group_gov_pol = {
 .attrs = dbs_attributes_gov_pol,
 .name = "ondemand",
};

/************************** sysfs end ************************/

static int od_init(struct dbs_data *dbs_data)
{
 struct od_dbs_tuners *tuners;
 u64 idle_time;
 int cpu;

 tuners = kzalloc(sizeof(struct od_dbs_tuners), GFP_KERNEL);
 if (!tuners) {
  pr_err("%s: kzalloc failed\n", __func__);
  return -ENOMEM;
 }

 cpu = get_cpu();
 idle_time = get_cpu_idle_time_us(cpu, NULL);
 put_cpu();
 if (idle_time != -1ULL) {
  /* Idle micro accounting is supported. Use finer thresholds */
  tuners->up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
  tuners->adj_up_threshold = MICRO_FREQUENCY_UP_THRESHOLD -
   MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
  /*
   * In nohz/micro accounting case we set the minimum frequency
   * not depending on HZ, but fixed (very low). The deferred
   * timer might skip some samples if idle/sleeping as needed.
  */
  dbs_data->min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
 } else {
  tuners->up_threshold = DEF_FREQUENCY_UP_THRESHOLD;
  tuners->adj_up_threshold = DEF_FREQUENCY_UP_THRESHOLD -
   DEF_FREQUENCY_DOWN_DIFFERENTIAL;

  /* For correct statistics, we need 10 ticks for each measure */
  dbs_data->min_sampling_rate = MIN_SAMPLING_RATE_RATIO *
   jiffies_to_usecs(10);
 }

 tuners->sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR;
 tuners->ignore_nice_load = 0;
 tuners->powersave_bias = default_powersave_bias;
 tuners->io_is_busy = should_io_be_busy();

 dbs_data->tuners = tuners;
 mutex_init(&dbs_data->mutex);
 return 0;
}

static void od_exit(struct dbs_data *dbs_data)
{
 kfree(dbs_data->tuners);
}

define_get_cpu_dbs_routines(od_cpu_dbs_info);

static struct od_ops od_ops = {
 .powersave_bias_init_cpu = ondemand_powersave_bias_init_cpu,
 .powersave_bias_target = generic_powersave_bias_target,
 .freq_increase = dbs_freq_increase,
};

static struct common_dbs_data od_dbs_cdata = {
 .governor = GOV_ONDEMAND,
 .attr_group_gov_sys = &od_attr_group_gov_sys,
 .attr_group_gov_pol = &od_attr_group_gov_pol,
 .get_cpu_cdbs = get_cpu_cdbs,
 .get_cpu_dbs_info_s = get_cpu_dbs_info_s,
 .gov_dbs_timer = od_dbs_timer,
 .gov_check_cpu = od_check_cpu,
 .gov_ops = &od_ops,
 .init = od_init,
 .exit = od_exit,
};

static void od_set_powersave_bias(unsigned int powersave_bias)
{
 struct cpufreq_policy *policy;
 struct dbs_data *dbs_data;
 struct od_dbs_tuners *od_tuners;
 unsigned int cpu;
 cpumask_t done;

 default_powersave_bias = powersave_bias;
 cpumask_clear(&done);

 get_online_cpus();
 for_each_online_cpu(cpu) {
  if (cpumask_test_cpu(cpu, &done))
   continue;

  policy = per_cpu(od_cpu_dbs_info, cpu).cdbs.cur_policy;
  if (!policy)
   continue;

  cpumask_or(&done, &done, policy->cpus);

  if (policy->governor != &cpufreq_gov_ondemand)
   continue;

  dbs_data = policy->governor_data;
  od_tuners = dbs_data->tuners;
  od_tuners->powersave_bias = default_powersave_bias;
 }
 put_online_cpus();
}

void od_register_powersave_bias_handler(unsigned int (*f)
  (struct cpufreq_policy *, unsigned int, unsigned int),
  unsigned int powersave_bias)
{
 od_ops.powersave_bias_target = f;
 od_set_powersave_bias(powersave_bias);
}
EXPORT_SYMBOL_GPL(od_register_powersave_bias_handler);

void od_unregister_powersave_bias_handler(void)
{
 od_ops.powersave_bias_target = generic_powersave_bias_target;
 od_set_powersave_bias(0);
}
EXPORT_SYMBOL_GPL(od_unregister_powersave_bias_handler);

static int od_cpufreq_governor_dbs(struct cpufreq_policy *policy,
  unsigned int event)
{
 return cpufreq_governor_dbs(policy, &od_dbs_cdata, event);
}

#ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
static
#endif
struct cpufreq_governor cpufreq_gov_ondemand = {
 .name   = "ondemand",
 .governor  = od_cpufreq_governor_dbs,
 .max_transition_latency = TRANSITION_LATENCY_LIMIT,
 .owner   = THIS_MODULE,
};

static int __init cpufreq_gov_dbs_init(void)
{
 return cpufreq_register_governor(&cpufreq_gov_ondemand);
}

static void __exit cpufreq_gov_dbs_exit(void)
{
 cpufreq_unregister_governor(&cpufreq_gov_ondemand);
}

MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 "Low Latency Frequency Transition capable processors");
MODULE_LICENSE("GPL");

#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
fs_initcall(cpufreq_gov_dbs_init);
#else
module_init(cpufreq_gov_dbs_init);
#endif
module_exit(cpufreq_gov_dbs_exit);