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