linux kernel -cpufreq_governor.c

/*
 * drivers/cpufreq/cpufreq_governor.c
 *
 * CPUFREQ governors common code
 *
 *
 * 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 <asm/cputime.h>
#include <linux/cpufreq.h>
#include <linux/cpumask.h>
#include <linux/export.h>
#include <linux/kernel_stat.h>
#include <linux/mutex.h>
#include <linux/slab.h>
#include <linux/types.h>
#include <linux/workqueue.h>

#include "cpufreq_governor.h"

static struct attribute_group *get_sysfs_attr(struct dbs_data *dbs_data)
{
 if (have_governor_per_policy())
  return dbs_data->cdata->attr_group_gov_pol;
 else
  return dbs_data->cdata->attr_group_gov_sys;
}

void dbs_check_cpu(struct dbs_data *dbs_data, int cpu)
{
 struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
 struct cpufreq_policy *policy;
 unsigned int max_load = 0;
 unsigned int ignore_nice;
 unsigned int j;

 if (dbs_data->cdata->governor == GOV_ONDEMAND)
  ignore_nice = od_tuners->ignore_nice_load;
 else
  ignore_nice = cs_tuners->ignore_nice_load;

 policy = cdbs->cur_policy;

 /* Get Absolute Load (in terms of freq for ondemand gov) */
 for_each_cpu(j, policy->cpus) {
  struct cpu_dbs_common_info *j_cdbs;
  u64 cur_wall_time, cur_idle_time;
  unsigned int idle_time, wall_time;
  unsigned int load;
  int io_busy = 0;

  j_cdbs = dbs_data->cdata->get_cpu_cdbs(j);

  /*
   * For the purpose of ondemand, waiting for disk IO is
   * an indication that you're performance critical, and
   * not that the system is actually idle. So do not add
   * the iowait time to the cpu idle time.
   */
  if (dbs_data->cdata->governor == GOV_ONDEMAND)
   io_busy = od_tuners->io_is_busy;
  cur_idle_time = get_cpu_idle_time(j, &cur_wall_time, io_busy);

  wall_time = (unsigned int)
   (cur_wall_time - j_cdbs->prev_cpu_wall);
  j_cdbs->prev_cpu_wall = cur_wall_time;

  idle_time = (unsigned int)
   (cur_idle_time - j_cdbs->prev_cpu_idle);
  j_cdbs->prev_cpu_idle = cur_idle_time;

  if (ignore_nice) {
   u64 cur_nice;
   unsigned long cur_nice_jiffies;

   cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] -
      cdbs->prev_cpu_nice;
   /*
    * Assumption: nice time between sampling periods will
    * be less than 2^32 jiffies for 32 bit sys
    */
   cur_nice_jiffies = (unsigned long)
     cputime64_to_jiffies64(cur_nice);

   cdbs->prev_cpu_nice =
    kcpustat_cpu(j).cpustat[CPUTIME_NICE];
   idle_time += jiffies_to_usecs(cur_nice_jiffies);
  }

  if (unlikely(!wall_time || wall_time < idle_time))
   continue;

  load = 100 * (wall_time - idle_time) / wall_time;

  if (dbs_data->cdata->governor == GOV_ONDEMAND) {
   int freq_avg = __cpufreq_driver_getavg(policy, j);
   if (freq_avg <= 0)
    freq_avg = policy->cur;

   load *= freq_avg;
  }

  if (load > max_load)
   max_load = load;
 }

 dbs_data->cdata->gov_check_cpu(cpu, max_load);
}
EXPORT_SYMBOL_GPL(dbs_check_cpu);

static inline void __gov_queue_work(int cpu, struct dbs_data *dbs_data,
  unsigned int delay)
{
 struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);

 mod_delayed_work_on(cpu, system_wq, &cdbs->work, delay);
}

void gov_queue_work(struct dbs_data *dbs_data, struct cpufreq_policy *policy,
  unsigned int delay, bool all_cpus)
{
 int i;

 if (!policy->governor_enabled)
  return;

 if (!all_cpus) {
  __gov_queue_work(smp_processor_id(), dbs_data, delay);
 } else {
  for_each_cpu(i, policy->cpus)
   __gov_queue_work(i, dbs_data, delay);
 }
}
EXPORT_SYMBOL_GPL(gov_queue_work);

static inline void gov_cancel_work(struct dbs_data *dbs_data,
  struct cpufreq_policy *policy)
{
 struct cpu_dbs_common_info *cdbs;
 int i;

 for_each_cpu(i, policy->cpus) {
  cdbs = dbs_data->cdata->get_cpu_cdbs(i);
  cancel_delayed_work_sync(&cdbs->work);
 }
}

/* Will return if we need to evaluate cpu load again or not */
bool need_load_eval(struct cpu_dbs_common_info *cdbs,
  unsigned int sampling_rate)
{
 if (policy_is_shared(cdbs->cur_policy)) {
  ktime_t time_now = ktime_get();
  s64 delta_us = ktime_us_delta(time_now, cdbs->time_stamp);

  /* Do nothing if we recently have sampled */
  if (delta_us < (s64)(sampling_rate / 2))
   return false;
  else
   cdbs->time_stamp = time_now;
 }

 return true;
}
EXPORT_SYMBOL_GPL(need_load_eval);

static void set_sampling_rate(struct dbs_data *dbs_data,
  unsigned int sampling_rate)
{
 if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
  struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
  cs_tuners->sampling_rate = sampling_rate;
 } else {
  struct od_dbs_tuners *od_tuners = dbs_data->tuners;
  od_tuners->sampling_rate = sampling_rate;
 }
}

int cpufreq_governor_dbs(struct cpufreq_policy *policy,
  struct common_dbs_data *cdata, unsigned int event)
{
 struct dbs_data *dbs_data;
 struct od_cpu_dbs_info_s *od_dbs_info = NULL;
 struct cs_cpu_dbs_info_s *cs_dbs_info = NULL;
 struct od_ops *od_ops = NULL;
 struct od_dbs_tuners *od_tuners = NULL;
 struct cs_dbs_tuners *cs_tuners = NULL;
 struct cpu_dbs_common_info *cpu_cdbs;
 unsigned int sampling_rate, latency, ignore_nice, j, cpu = policy->cpu;
 int io_busy = 0;
 int rc;

 if (have_governor_per_policy())
  dbs_data = policy->governor_data;
 else
  dbs_data = cdata->gdbs_data;

 WARN_ON(!dbs_data && (event != CPUFREQ_GOV_POLICY_INIT));

 switch (event) {
 case CPUFREQ_GOV_POLICY_INIT:
  if (have_governor_per_policy()) {
   WARN_ON(dbs_data);
  } else if (dbs_data) {
   dbs_data->usage_count++;
   policy->governor_data = dbs_data;
   return 0;
  }

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

  dbs_data->cdata = cdata;
  dbs_data->usage_count = 1;
  rc = cdata->init(dbs_data);
  if (rc) {
   pr_err("%s: POLICY_INIT: init() failed\n", __func__);
   kfree(dbs_data);
   return rc;
  }

  rc = sysfs_create_group(get_governor_parent_kobj(policy),
    get_sysfs_attr(dbs_data));
  if (rc) {
   cdata->exit(dbs_data);
   kfree(dbs_data);
   return rc;
  }

  policy->governor_data = dbs_data;

  /* policy latency is in nS. Convert it to uS first */
  latency = policy->cpuinfo.transition_latency / 1000;
  if (latency == 0)
   latency = 1;

  /* Bring kernel and HW constraints together */
  dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate,
    MIN_LATENCY_MULTIPLIER * latency);
  set_sampling_rate(dbs_data, max(dbs_data->min_sampling_rate,
     latency * LATENCY_MULTIPLIER));

  if ((cdata->governor == GOV_CONSERVATIVE) &&
    (!policy->governor->initialized)) {
   struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;

   cpufreq_register_notifier(cs_ops->notifier_block,
     CPUFREQ_TRANSITION_NOTIFIER);
  }

  if (!have_governor_per_policy())
   cdata->gdbs_data = dbs_data;

  return 0;
 case CPUFREQ_GOV_POLICY_EXIT:
  if (!--dbs_data->usage_count) {
   sysfs_remove_group(get_governor_parent_kobj(policy),
     get_sysfs_attr(dbs_data));

   if ((dbs_data->cdata->governor == GOV_CONSERVATIVE) &&
    (policy->governor->initialized == 1)) {
    struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;

    cpufreq_unregister_notifier(cs_ops->notifier_block,
      CPUFREQ_TRANSITION_NOTIFIER);
   }

   cdata->exit(dbs_data);
   kfree(dbs_data);
   cdata->gdbs_data = NULL;
  }

  policy->governor_data = NULL;
  return 0;
 }

 cpu_cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);

 if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
  cs_tuners = dbs_data->tuners;
  cs_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
  sampling_rate = cs_tuners->sampling_rate;
  ignore_nice = cs_tuners->ignore_nice_load;
 } else {
  od_tuners = dbs_data->tuners;
  od_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
  sampling_rate = od_tuners->sampling_rate;
  ignore_nice = od_tuners->ignore_nice_load;
  od_ops = dbs_data->cdata->gov_ops;
  io_busy = od_tuners->io_is_busy;
 }

 switch (event) {
 case CPUFREQ_GOV_START:
  if (!policy->cur)
   return -EINVAL;

  mutex_lock(&dbs_data->mutex);

  for_each_cpu(j, policy->cpus) {
   struct cpu_dbs_common_info *j_cdbs =
    dbs_data->cdata->get_cpu_cdbs(j);

   j_cdbs->cpu = j;
   j_cdbs->cur_policy = policy;
   j_cdbs->prev_cpu_idle = get_cpu_idle_time(j,
            &j_cdbs->prev_cpu_wall, io_busy);
   if (ignore_nice)
    j_cdbs->prev_cpu_nice =
     kcpustat_cpu(j).cpustat[CPUTIME_NICE];

   mutex_init(&j_cdbs->timer_mutex);
   INIT_DEFERRABLE_WORK(&j_cdbs->work,
          dbs_data->cdata->gov_dbs_timer);
  }

  /*
   * conservative does not implement micro like ondemand
   * governor, thus we are bound to jiffes/HZ
   */
  if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
   cs_dbs_info->down_skip = 0;
   cs_dbs_info->enable = 1;
   cs_dbs_info->requested_freq = policy->cur;
  } else {
   od_dbs_info->rate_mult = 1;
   od_dbs_info->sample_type = OD_NORMAL_SAMPLE;
   od_ops->powersave_bias_init_cpu(cpu);
  }

  mutex_unlock(&dbs_data->mutex);

  /* Initiate timer time stamp */
  cpu_cdbs->time_stamp = ktime_get();

  gov_queue_work(dbs_data, policy,
    delay_for_sampling_rate(sampling_rate), true);
  break;

 case CPUFREQ_GOV_STOP:
  if (dbs_data->cdata->governor == GOV_CONSERVATIVE)
   cs_dbs_info->enable = 0;

  gov_cancel_work(dbs_data, policy);

  mutex_lock(&dbs_data->mutex);
  mutex_destroy(&cpu_cdbs->timer_mutex);

  mutex_unlock(&dbs_data->mutex);

  break;

 case CPUFREQ_GOV_LIMITS:
  mutex_lock(&cpu_cdbs->timer_mutex);
  if (policy->max < cpu_cdbs->cur_policy->cur)
   __cpufreq_driver_target(cpu_cdbs->cur_policy,
     policy->max, CPUFREQ_RELATION_H);
  else if (policy->min > cpu_cdbs->cur_policy->cur)
   __cpufreq_driver_target(cpu_cdbs->cur_policy,
     policy->min, CPUFREQ_RELATION_L);
  dbs_check_cpu(dbs_data, cpu);
  mutex_unlock(&cpu_cdbs->timer_mutex);
  break;
 }
 return 0;
}
EXPORT_SYMBOL_GPL(cpufreq_governor_dbs);