drivers/cpufreq/cpufreq_governor.c

   1 /*
   2  * drivers/cpufreq/cpufreq_governor.c
   3  *
   4  * CPUFREQ governors common code
   5  *
   6  * Copyright    (C) 2001 Russell King
   7  *              (C) 2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
   8  *              (C) 2003 Jun Nakajima <jun.nakajima@intel.com>
   9  *              (C) 2009 Alexander Clouter <alex@digriz.org.uk>
  10  *              (c) 2012 Viresh Kumar <viresh.kumar@linaro.org>
  11  *
  12  * This program is free software; you can redistribute it and/or modify
  13  * it under the terms of the GNU General Public License version 2 as
  14  * published by the Free Software Foundation.
  15  */
  16
  17 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  18
  19 #include <asm/cputime.h>
  20 #include <linux/cpufreq.h>
  21 #include <linux/cpumask.h>
  22 #include <linux/export.h>
  23 #include <linux/kernel_stat.h>
  24 #include <linux/mutex.h>
  25 #include <linux/slab.h>
  26 #include <linux/types.h>
  27 #include <linux/workqueue.h>
  28
  29 #include "cpufreq_governor.h"
  30 #include "cpu_load_metric.h"
  31
  32 static struct attribute_group *get_sysfs_attr(struct dbs_data *dbs_data)
  33 {
  34         if (have_governor_per_policy())
  35                 return dbs_data->cdata->attr_group_gov_pol;
  36         else
  37                 return dbs_data->cdata->attr_group_gov_sys;
  38 }
  39
  40 void dbs_check_cpu(struct dbs_data *dbs_data, int cpu)
  41 {
  42         struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
  43         struct od_dbs_tuners *od_tuners = dbs_data->tuners;
  44         struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
  45         struct cpufreq_policy *policy;
  46         unsigned int max_load = 0;
  47         unsigned int ignore_nice;
  48         unsigned int j;
  49
  50         if (dbs_data->cdata->governor == GOV_ONDEMAND)
  51                 ignore_nice = od_tuners->ignore_nice_load;
  52         else
  53                 ignore_nice = cs_tuners->ignore_nice_load;
  54
  55         policy = cdbs->cur_policy;
  56
  57         /* Get Absolute Load */
  58         for_each_cpu(j, policy->cpus) {
  59                 struct cpu_dbs_common_info *j_cdbs;
  60                 u64 cur_wall_time, cur_idle_time;
  61                 unsigned int idle_time, wall_time;
  62                 unsigned int load;
  63                 int io_busy = 0;
  64
  65                 j_cdbs = dbs_data->cdata->get_cpu_cdbs(j);
  66
  67                 /*
  68                  * For the purpose of ondemand, waiting for disk IO is
  69                  * an indication that you're performance critical, and
  70                  * not that the system is actually idle. So do not add
  71                  * the iowait time to the cpu idle time.
  72                  */
  73                 if (dbs_data->cdata->governor == GOV_ONDEMAND)
  74                         io_busy = od_tuners->io_is_busy;
  75                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time, io_busy);
  76
  77                 wall_time = (unsigned int)
  78                         (cur_wall_time - j_cdbs->prev_cpu_wall);
  79                 j_cdbs->prev_cpu_wall = cur_wall_time;
  80
  81                 idle_time = (unsigned int)
  82                         (cur_idle_time - j_cdbs->prev_cpu_idle);
  83                 j_cdbs->prev_cpu_idle = cur_idle_time;
  84
  85                 if (ignore_nice) {
  86                         u64 cur_nice;
  87                         unsigned long cur_nice_jiffies;
  88
  89                         cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] -
  90                                          cdbs->prev_cpu_nice;
  91                         /*
  92                          * Assumption: nice time between sampling periods will
  93                          * be less than 2^32 jiffies for 32 bit sys
  94                          */
  95                         cur_nice_jiffies = (unsigned long)
  96                                         cputime64_to_jiffies64(cur_nice);
  97
  98                         cdbs->prev_cpu_nice =
  99                                 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 100                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
 101                 }
 102
 103                 if (unlikely(!wall_time || wall_time < idle_time))
 104                         continue;
 105
 106                 load = 100 * (wall_time - idle_time) / wall_time;
 107
 108                 if (load > max_load)
 109                         max_load = load;
 110
 111                 update_cpu_metric(j, cur_wall_time, idle_time, wall_time, policy);
 112         }
 113
 114         dbs_data->cdata->gov_check_cpu(cpu, max_load);
 115 }
 116 EXPORT_SYMBOL_GPL(dbs_check_cpu);
 117
 118 static inline void __gov_queue_work(int cpu, struct dbs_data *dbs_data,
 119                 unsigned int delay)
 120 {
 121         struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
 122
 123         mod_delayed_work_on(cpu, system_wq, &cdbs->work, delay);
 124 }
 125
 126 void gov_queue_work(struct dbs_data *dbs_data, struct cpufreq_policy *policy,
 127                 unsigned int delay, bool all_cpus)
 128 {
 129         int i;
 130
 131         if (!policy->governor_enabled)
 132                 return;
 133
 134         if (!all_cpus) {
 135                 __gov_queue_work(smp_processor_id(), dbs_data, delay);
 136         } else {
 137                 for_each_cpu(i, policy->cpus)
 138                         __gov_queue_work(i, dbs_data, delay);
 139         }
 140 }
 141 EXPORT_SYMBOL_GPL(gov_queue_work);
 142
 143 static inline void gov_cancel_work(struct dbs_data *dbs_data,
 144                 struct cpufreq_policy *policy)
 145 {
 146         struct cpu_dbs_common_info *cdbs;
 147         int i;
 148
 149         for_each_cpu(i, policy->cpus) {
 150                 cdbs = dbs_data->cdata->get_cpu_cdbs(i);
 151                 cancel_delayed_work_sync(&cdbs->work);
 152         }
 153 }
 154
 155 /* Will return if we need to evaluate cpu load again or not */
 156 bool need_load_eval(struct cpu_dbs_common_info *cdbs,
 157                 unsigned int sampling_rate)
 158 {
 159         if (policy_is_shared(cdbs->cur_policy)) {
 160                 ktime_t time_now = ktime_get();
 161                 s64 delta_us = ktime_us_delta(time_now, cdbs->time_stamp);
 162
 163                 /* Do nothing if we recently have sampled */
 164                 if (delta_us < (s64)(sampling_rate / 2))
 165                         return false;
 166                 else
 167                         cdbs->time_stamp = time_now;
 168         }
 169
 170         return true;
 171 }
 172 EXPORT_SYMBOL_GPL(need_load_eval);
 173
 174 static void set_sampling_rate(struct dbs_data *dbs_data,
 175                 unsigned int sampling_rate)
 176 {
 177         if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
 178                 struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
 179                 cs_tuners->sampling_rate = sampling_rate;
 180         } else {
 181                 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 182                 od_tuners->sampling_rate = sampling_rate;
 183         }
 184 }
 185
 186 int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 187                 struct common_dbs_data *cdata, unsigned int event)
 188 {
 189         struct dbs_data *dbs_data;
 190         struct od_cpu_dbs_info_s *od_dbs_info = NULL;
 191         struct cs_cpu_dbs_info_s *cs_dbs_info = NULL;
 192         struct od_ops *od_ops = NULL;
 193         struct od_dbs_tuners *od_tuners = NULL;
 194         struct cs_dbs_tuners *cs_tuners = NULL;
 195         struct cpu_dbs_common_info *cpu_cdbs;
 196         unsigned int sampling_rate, latency, ignore_nice, j, cpu = policy->cpu;
 197         int io_busy = 0;
 198         int rc;
 199
 200         if (have_governor_per_policy())
 201                 dbs_data = policy->governor_data;
 202         else
 203                 dbs_data = cdata->gdbs_data;
 204
 205         WARN_ON(!dbs_data && (event != CPUFREQ_GOV_POLICY_INIT));
 206
 207         switch (event) {
 208         case CPUFREQ_GOV_POLICY_INIT:
 209                 if (have_governor_per_policy()) {
 210                         WARN_ON(dbs_data);
 211                 } else if (dbs_data) {
 212                         dbs_data->usage_count++;
 213                         policy->governor_data = dbs_data;
 214                         return 0;
 215                 }
 216
 217                 dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
 218                 if (!dbs_data) {
 219                         pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
 220                         return -ENOMEM;
 221                 }
 222
 223                 dbs_data->cdata = cdata;
 224                 dbs_data->usage_count = 1;
 225                 rc = cdata->init(dbs_data);
 226                 if (rc) {
 227                         pr_err("%s: POLICY_INIT: init() failed\n", __func__);
 228                         kfree(dbs_data);
 229                         return rc;
 230                 }
 231
 232                 rc = sysfs_create_group(get_governor_parent_kobj(policy),
 233                                 get_sysfs_attr(dbs_data));
 234                 if (rc) {
 235                         cdata->exit(dbs_data);
 236                         kfree(dbs_data);
 237                         return rc;
 238                 }
 239
 240                 policy->governor_data = dbs_data;
 241
 242                 /* policy latency is in nS. Convert it to uS first */
 243                 latency = policy->cpuinfo.transition_latency / 1000;
 244                 if (latency == 0)
 245                         latency = 1;
 246
 247                 /* Bring kernel and HW constraints together */
 248                 dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate,
 249                                 MIN_LATENCY_MULTIPLIER * latency);
 250                 set_sampling_rate(dbs_data, max(dbs_data->min_sampling_rate,
 251                                         latency * LATENCY_MULTIPLIER));
 252
 253                 if ((cdata->governor == GOV_CONSERVATIVE) &&
 254                                 (!policy->governor->initialized)) {
 255                         struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
 256
 257                         cpufreq_register_notifier(cs_ops->notifier_block,
 258                                         CPUFREQ_TRANSITION_NOTIFIER);
 259                 }
 260
 261                 if (!have_governor_per_policy())
 262                         cdata->gdbs_data = dbs_data;
 263
 264                 return 0;
 265         case CPUFREQ_GOV_POLICY_EXIT:
 266                 if (!--dbs_data->usage_count) {
 267                         sysfs_remove_group(get_governor_parent_kobj(policy),
 268                                         get_sysfs_attr(dbs_data));
 269
 270                         if ((dbs_data->cdata->governor == GOV_CONSERVATIVE) &&
 271                                 (policy->governor->initialized == 1)) {
 272                                 struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
 273
 274                                 cpufreq_unregister_notifier(cs_ops->notifier_block,
 275                                                 CPUFREQ_TRANSITION_NOTIFIER);
 276                         }
 277
 278                         cdata->exit(dbs_data);
 279                         kfree(dbs_data);
 280                         cdata->gdbs_data = NULL;
 281                 }
 282
 283                 policy->governor_data = NULL;
 284                 return 0;
 285         }
 286
 287         cpu_cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
 288
 289         if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
 290                 cs_tuners = dbs_data->tuners;
 291                 cs_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
 292                 sampling_rate = cs_tuners->sampling_rate;
 293                 ignore_nice = cs_tuners->ignore_nice_load;
 294         } else {
 295                 od_tuners = dbs_data->tuners;
 296                 od_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
 297                 sampling_rate = od_tuners->sampling_rate;
 298                 ignore_nice = od_tuners->ignore_nice_load;
 299                 od_ops = dbs_data->cdata->gov_ops;
 300                 io_busy = od_tuners->io_is_busy;
 301         }
 302
 303         switch (event) {
 304         case CPUFREQ_GOV_START:
 305                 if (!policy->cur)
 306                         return -EINVAL;
 307
 308                 mutex_lock(&dbs_data->mutex);
 309
 310                 for_each_cpu(j, policy->cpus) {
 311                         struct cpu_dbs_common_info *j_cdbs =
 312                                 dbs_data->cdata->get_cpu_cdbs(j);
 313
 314                         j_cdbs->cpu = j;
 315                         j_cdbs->cur_policy = policy;
 316                         j_cdbs->prev_cpu_idle = get_cpu_idle_time(j,
 317                                                &j_cdbs->prev_cpu_wall, io_busy);
 318                         if (ignore_nice)
 319                                 j_cdbs->prev_cpu_nice =
 320                                         kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 321
 322                         mutex_init(&j_cdbs->timer_mutex);
 323                         INIT_DEFERRABLE_WORK(&j_cdbs->work,
 324                                              dbs_data->cdata->gov_dbs_timer);
 325                 }
 326
 327                 /*
 328                  * conservative does not implement micro like ondemand
 329                  * governor, thus we are bound to jiffes/HZ
 330                  */
 331                 if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
 332                         cs_dbs_info->down_skip = 0;
 333                         cs_dbs_info->enable = 1;
 334                         cs_dbs_info->requested_freq = policy->cur;
 335                 } else {
 336                         od_dbs_info->rate_mult = 1;
 337                         od_dbs_info->sample_type = OD_NORMAL_SAMPLE;
 338                         od_ops->powersave_bias_init_cpu(cpu);
 339                 }
 340
 341                 mutex_unlock(&dbs_data->mutex);
 342
 343                 /* Initiate timer time stamp */
 344                 cpu_cdbs->time_stamp = ktime_get();
 345
 346                 gov_queue_work(dbs_data, policy,
 347                                 delay_for_sampling_rate(sampling_rate), true);
 348                 break;
 349
 350         case CPUFREQ_GOV_STOP:
 351                 if (dbs_data->cdata->governor == GOV_CONSERVATIVE)
 352                         cs_dbs_info->enable = 0;
 353
 354                 gov_cancel_work(dbs_data, policy);
 355
 356                 mutex_lock(&dbs_data->mutex);
 357                 mutex_destroy(&cpu_cdbs->timer_mutex);
 358                 cpu_cdbs->cur_policy = NULL;
 359
 360                 mutex_unlock(&dbs_data->mutex);
 361
 362                 break;
 363
 364         case CPUFREQ_GOV_LIMITS:
 365                 mutex_lock(&cpu_cdbs->timer_mutex);
 366                 if (policy->max < cpu_cdbs->cur_policy->cur)
 367                         __cpufreq_driver_target(cpu_cdbs->cur_policy,
 368                                         policy->max, CPUFREQ_RELATION_H);
 369                 else if (policy->min > cpu_cdbs->cur_policy->cur)
 370                         __cpufreq_driver_target(cpu_cdbs->cur_policy,
 371                                         policy->min, CPUFREQ_RELATION_L);
 372                 dbs_check_cpu(dbs_data, cpu);
 373                 mutex_unlock(&cpu_cdbs->timer_mutex);
 374                 break;
 375         }
 376         return 0;
 377 }
 378 EXPORT_SYMBOL_GPL(cpufreq_governor_dbs);