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/tick.h>
  27 #include <linux/types.h>
  28 #include <linux/workqueue.h>
  29
  30 #include "cpufreq_governor.h"
  31
  32 static struct kobject *get_governor_parent_kobj(struct cpufreq_policy *policy)
  33 {
  34         if (have_governor_per_policy())
  35                 return &policy->kobj;
  36         else
  37                 return cpufreq_global_kobject;
  38 }
  39
  40 static struct attribute_group *get_sysfs_attr(struct dbs_data *dbs_data)
  41 {
  42         if (have_governor_per_policy())
  43                 return dbs_data->cdata->attr_group_gov_pol;
  44         else
  45                 return dbs_data->cdata->attr_group_gov_sys;
  46 }
  47
  48 static inline u64 get_cpu_idle_time_jiffy(unsigned int cpu, u64 *wall)
  49 {
  50         u64 idle_time;
  51         u64 cur_wall_time;
  52         u64 busy_time;
  53
  54         cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
  55
  56         busy_time = kcpustat_cpu(cpu).cpustat[CPUTIME_USER];
  57         busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_SYSTEM];
  58         busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_IRQ];
  59         busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_SOFTIRQ];
  60         busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_STEAL];
  61         busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_NICE];
  62
  63         idle_time = cur_wall_time - busy_time;
  64         if (wall)
  65                 *wall = cputime_to_usecs(cur_wall_time);
  66
  67         return cputime_to_usecs(idle_time);
  68 }
  69
  70 u64 get_cpu_idle_time(unsigned int cpu, u64 *wall)
  71 {
  72         u64 idle_time = get_cpu_idle_time_us(cpu, NULL);
  73
  74         if (idle_time == -1ULL)
  75                 return get_cpu_idle_time_jiffy(cpu, wall);
  76         else
  77                 idle_time += get_cpu_iowait_time_us(cpu, wall);
  78
  79         return idle_time;
  80 }
  81 EXPORT_SYMBOL_GPL(get_cpu_idle_time);
  82
  83 void dbs_check_cpu(struct dbs_data *dbs_data, int cpu)
  84 {
  85         struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
  86         struct od_dbs_tuners *od_tuners = dbs_data->tuners;
  87         struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
  88         struct cpufreq_policy *policy;
  89         unsigned int max_load = 0;
  90         unsigned int ignore_nice;
  91         unsigned int j;
  92
  93         if (dbs_data->cdata->governor == GOV_ONDEMAND)
  94                 ignore_nice = od_tuners->ignore_nice;
  95         else
  96                 ignore_nice = cs_tuners->ignore_nice;
  97
  98         policy = cdbs->cur_policy;
  99
 100         /* Get Absolute Load (in terms of freq for ondemand gov) */
 101         for_each_cpu(j, policy->cpus) {
 102                 struct cpu_dbs_common_info *j_cdbs;
 103                 u64 cur_wall_time, cur_idle_time, cur_iowait_time;
 104                 unsigned int idle_time, wall_time, iowait_time;
 105                 unsigned int load;
 106
 107                 j_cdbs = dbs_data->cdata->get_cpu_cdbs(j);
 108
 109                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
 110
 111                 wall_time = (unsigned int)
 112                         (cur_wall_time - j_cdbs->prev_cpu_wall);
 113                 j_cdbs->prev_cpu_wall = cur_wall_time;
 114
 115                 idle_time = (unsigned int)
 116                         (cur_idle_time - j_cdbs->prev_cpu_idle);
 117                 j_cdbs->prev_cpu_idle = cur_idle_time;
 118
 119                 if (ignore_nice) {
 120                         u64 cur_nice;
 121                         unsigned long cur_nice_jiffies;
 122
 123                         cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] -
 124                                          cdbs->prev_cpu_nice;
 125                         /*
 126                          * Assumption: nice time between sampling periods will
 127                          * be less than 2^32 jiffies for 32 bit sys
 128                          */
 129                         cur_nice_jiffies = (unsigned long)
 130                                         cputime64_to_jiffies64(cur_nice);
 131
 132                         cdbs->prev_cpu_nice =
 133                                 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 134                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
 135                 }
 136
 137                 if (dbs_data->cdata->governor == GOV_ONDEMAND) {
 138                         struct od_cpu_dbs_info_s *od_j_dbs_info =
 139                                 dbs_data->cdata->get_cpu_dbs_info_s(cpu);
 140
 141                         cur_iowait_time = get_cpu_iowait_time_us(j,
 142                                         &cur_wall_time);
 143                         if (cur_iowait_time == -1ULL)
 144                                 cur_iowait_time = 0;
 145
 146                         iowait_time = (unsigned int) (cur_iowait_time -
 147                                         od_j_dbs_info->prev_cpu_iowait);
 148                         od_j_dbs_info->prev_cpu_iowait = cur_iowait_time;
 149
 150                         /*
 151                          * For the purpose of ondemand, waiting for disk IO is
 152                          * an indication that you're performance critical, and
 153                          * not that the system is actually idle. So subtract the
 154                          * iowait time from the cpu idle time.
 155                          */
 156                         if (od_tuners->io_is_busy && idle_time >= iowait_time)
 157                                 idle_time -= iowait_time;
 158                 }
 159
 160                 if (unlikely(!wall_time || wall_time < idle_time))
 161                         continue;
 162
 163                 load = 100 * (wall_time - idle_time) / wall_time;
 164
 165                 if (dbs_data->cdata->governor == GOV_ONDEMAND) {
 166                         int freq_avg = __cpufreq_driver_getavg(policy, j);
 167                         if (freq_avg <= 0)
 168                                 freq_avg = policy->cur;
 169
 170                         load *= freq_avg;
 171                 }
 172
 173                 if (load > max_load)
 174                         max_load = load;
 175         }
 176
 177         dbs_data->cdata->gov_check_cpu(cpu, max_load);
 178 }
 179 EXPORT_SYMBOL_GPL(dbs_check_cpu);
 180
 181 static inline void dbs_timer_init(struct dbs_data *dbs_data, int cpu,
 182                                   unsigned int sampling_rate)
 183 {
 184         int delay = delay_for_sampling_rate(sampling_rate);
 185         struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
 186
 187         schedule_delayed_work_on(cpu, &cdbs->work, delay);
 188 }
 189
 190 static inline void dbs_timer_exit(struct dbs_data *dbs_data, int cpu)
 191 {
 192         struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
 193
 194         cancel_delayed_work_sync(&cdbs->work);
 195 }
 196
 197 /* Will return if we need to evaluate cpu load again or not */
 198 bool need_load_eval(struct cpu_dbs_common_info *cdbs,
 199                 unsigned int sampling_rate)
 200 {
 201         if (policy_is_shared(cdbs->cur_policy)) {
 202                 ktime_t time_now = ktime_get();
 203                 s64 delta_us = ktime_us_delta(time_now, cdbs->time_stamp);
 204
 205                 /* Do nothing if we recently have sampled */
 206                 if (delta_us < (s64)(sampling_rate / 2))
 207                         return false;
 208                 else
 209                         cdbs->time_stamp = time_now;
 210         }
 211
 212         return true;
 213 }
 214 EXPORT_SYMBOL_GPL(need_load_eval);
 215
 216 static void set_sampling_rate(struct dbs_data *dbs_data,
 217                 unsigned int sampling_rate)
 218 {
 219         if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
 220                 struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
 221                 cs_tuners->sampling_rate = sampling_rate;
 222         } else {
 223                 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 224                 od_tuners->sampling_rate = sampling_rate;
 225         }
 226 }
 227
 228 int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 229                 struct common_dbs_data *cdata, unsigned int event)
 230 {
 231         struct dbs_data *dbs_data;
 232         struct od_cpu_dbs_info_s *od_dbs_info = NULL;
 233         struct cs_cpu_dbs_info_s *cs_dbs_info = NULL;
 234         struct od_ops *od_ops = NULL;
 235         struct od_dbs_tuners *od_tuners = NULL;
 236         struct cs_dbs_tuners *cs_tuners = NULL;
 237         struct cpu_dbs_common_info *cpu_cdbs;
 238         unsigned int sampling_rate, latency, ignore_nice, j, cpu = policy->cpu;
 239         int rc;
 240
 241         if (have_governor_per_policy())
 242                 dbs_data = policy->governor_data;
 243         else
 244                 dbs_data = cdata->gdbs_data;
 245
 246         WARN_ON(!dbs_data && (event != CPUFREQ_GOV_POLICY_INIT));
 247
 248         switch (event) {
 249         case CPUFREQ_GOV_POLICY_INIT:
 250                 if (have_governor_per_policy()) {
 251                         WARN_ON(dbs_data);
 252                 } else if (dbs_data) {
 253                         policy->governor_data = dbs_data;
 254                         return 0;
 255                 }
 256
 257                 dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
 258                 if (!dbs_data) {
 259                         pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
 260                         return -ENOMEM;
 261                 }
 262
 263                 dbs_data->cdata = cdata;
 264                 rc = cdata->init(dbs_data);
 265                 if (rc) {
 266                         pr_err("%s: POLICY_INIT: init() failed\n", __func__);
 267                         kfree(dbs_data);
 268                         return rc;
 269                 }
 270
 271                 rc = sysfs_create_group(get_governor_parent_kobj(policy),
 272                                 get_sysfs_attr(dbs_data));
 273                 if (rc) {
 274                         cdata->exit(dbs_data);
 275                         kfree(dbs_data);
 276                         return rc;
 277                 }
 278
 279                 policy->governor_data = dbs_data;
 280
 281                 /* policy latency is in nS. Convert it to uS first */
 282                 latency = policy->cpuinfo.transition_latency / 1000;
 283                 if (latency == 0)
 284                         latency = 1;
 285
 286                 /* Bring kernel and HW constraints together */
 287                 dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate,
 288                                 MIN_LATENCY_MULTIPLIER * latency);
 289                 set_sampling_rate(dbs_data, max(dbs_data->min_sampling_rate,
 290                                         latency * LATENCY_MULTIPLIER));
 291
 292                 if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
 293                         struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
 294
 295                         cpufreq_register_notifier(cs_ops->notifier_block,
 296                                         CPUFREQ_TRANSITION_NOTIFIER);
 297                 }
 298
 299                 if (!have_governor_per_policy())
 300                         cdata->gdbs_data = dbs_data;
 301
 302                 return 0;
 303         case CPUFREQ_GOV_POLICY_EXIT:
 304                 if ((policy->governor->initialized == 1) ||
 305                                 have_governor_per_policy()) {
 306                         sysfs_remove_group(get_governor_parent_kobj(policy),
 307                                         get_sysfs_attr(dbs_data));
 308
 309                         if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
 310                                 struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
 311
 312                                 cpufreq_unregister_notifier(cs_ops->notifier_block,
 313                                                 CPUFREQ_TRANSITION_NOTIFIER);
 314                         }
 315
 316                         cdata->exit(dbs_data);
 317                         kfree(dbs_data);
 318                         cdata->gdbs_data = NULL;
 319                 }
 320
 321                 policy->governor_data = NULL;
 322                 return 0;
 323         }
 324
 325         cpu_cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
 326
 327         if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
 328                 cs_tuners = dbs_data->tuners;
 329                 cs_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
 330                 sampling_rate = cs_tuners->sampling_rate;
 331                 ignore_nice = cs_tuners->ignore_nice;
 332         } else {
 333                 od_tuners = dbs_data->tuners;
 334                 od_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
 335                 sampling_rate = od_tuners->sampling_rate;
 336                 ignore_nice = od_tuners->ignore_nice;
 337                 od_ops = dbs_data->cdata->gov_ops;
 338         }
 339
 340         switch (event) {
 341         case CPUFREQ_GOV_START:
 342                 if (!policy->cur)
 343                         return -EINVAL;
 344
 345                 mutex_lock(&dbs_data->mutex);
 346
 347                 for_each_cpu(j, policy->cpus) {
 348                         struct cpu_dbs_common_info *j_cdbs =
 349                                 dbs_data->cdata->get_cpu_cdbs(j);
 350
 351                         j_cdbs->cpu = j;
 352                         j_cdbs->cur_policy = policy;
 353                         j_cdbs->prev_cpu_idle = get_cpu_idle_time(j,
 354                                         &j_cdbs->prev_cpu_wall);
 355                         if (ignore_nice)
 356                                 j_cdbs->prev_cpu_nice =
 357                                         kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 358
 359                         mutex_init(&j_cdbs->timer_mutex);
 360                         INIT_DEFERRABLE_WORK(&j_cdbs->work,
 361                                              dbs_data->cdata->gov_dbs_timer);
 362                 }
 363
 364                 /*
 365                  * conservative does not implement micro like ondemand
 366                  * governor, thus we are bound to jiffes/HZ
 367                  */
 368                 if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
 369                         cs_dbs_info->down_skip = 0;
 370                         cs_dbs_info->enable = 1;
 371                         cs_dbs_info->requested_freq = policy->cur;
 372                 } else {
 373                         od_dbs_info->rate_mult = 1;
 374                         od_dbs_info->sample_type = OD_NORMAL_SAMPLE;
 375                         od_ops->powersave_bias_init_cpu(cpu);
 376                 }
 377
 378                 mutex_unlock(&dbs_data->mutex);
 379
 380                 /* Initiate timer time stamp */
 381                 cpu_cdbs->time_stamp = ktime_get();
 382
 383                 for_each_cpu(j, policy->cpus)
 384                         dbs_timer_init(dbs_data, j, sampling_rate);
 385                 break;
 386
 387         case CPUFREQ_GOV_STOP:
 388                 if (dbs_data->cdata->governor == GOV_CONSERVATIVE)
 389                         cs_dbs_info->enable = 0;
 390
 391                 for_each_cpu(j, policy->cpus)
 392                         dbs_timer_exit(dbs_data, j);
 393
 394                 mutex_lock(&dbs_data->mutex);
 395                 mutex_destroy(&cpu_cdbs->timer_mutex);
 396
 397                 mutex_unlock(&dbs_data->mutex);
 398
 399                 break;
 400
 401         case CPUFREQ_GOV_LIMITS:
 402                 mutex_lock(&cpu_cdbs->timer_mutex);
 403                 if (policy->max < cpu_cdbs->cur_policy->cur)
 404                         __cpufreq_driver_target(cpu_cdbs->cur_policy,
 405                                         policy->max, CPUFREQ_RELATION_H);
 406                 else if (policy->min > cpu_cdbs->cur_policy->cur)
 407                         __cpufreq_driver_target(cpu_cdbs->cur_policy,
 408                                         policy->min, CPUFREQ_RELATION_L);
 409                 dbs_check_cpu(dbs_data, cpu);
 410                 mutex_unlock(&cpu_cdbs->timer_mutex);
 411                 break;
 412         }
 413         return 0;
 414 }
 415 EXPORT_SYMBOL_GPL(cpufreq_governor_dbs);