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  * Copyright (c) 2014, NVIDIA CORPORATION.  All rights reserved.
  12  *
  13  * This program is free software; you can redistribute it and/or modify
  14  * it under the terms of the GNU General Public License version 2 as
  15  * published by the Free Software Foundation.
  16  */
  17
  18 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  19
  20 #include <linux/export.h>
  21 #include <linux/kernel_stat.h>
  22 #include <linux/slab.h>
  23
  24 #include "cpufreq_governor.h"
  25
  26 static struct attribute_group *get_sysfs_attr(struct dbs_data *dbs_data)
  27 {
  28         if (have_governor_per_policy())
  29                 return dbs_data->cdata->attr_group_gov_pol;
  30         else
  31                 return dbs_data->cdata->attr_group_gov_sys;
  32 }
  33
  34 void dbs_check_cpu(struct dbs_data *dbs_data, int cpu)
  35 {
  36         struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
  37         struct od_dbs_tuners *od_tuners = dbs_data->tuners;
  38         struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
  39         struct cpufreq_policy *policy;
  40         unsigned int max_load = 0;
  41         unsigned int ignore_nice;
  42         unsigned int j;
  43
  44         if (dbs_data->cdata->governor == GOV_ONDEMAND)
  45                 ignore_nice = od_tuners->ignore_nice_load;
  46         else
  47                 ignore_nice = cs_tuners->ignore_nice_load;
  48
  49         policy = cdbs->cur_policy;
  50
  51         /* Get Absolute Load */
  52         for_each_cpu(j, policy->cpus) {
  53                 struct cpu_dbs_common_info *j_cdbs;
  54                 u64 cur_wall_time, cur_idle_time;
  55                 unsigned int idle_time, wall_time;
  56                 unsigned int load;
  57                 int io_busy = 0;
  58
  59                 j_cdbs = dbs_data->cdata->get_cpu_cdbs(j);
  60
  61                 /*
  62                  * For the purpose of ondemand, waiting for disk IO is
  63                  * an indication that you're performance critical, and
  64                  * not that the system is actually idle. So do not add
  65                  * the iowait time to the cpu idle time.
  66                  */
  67                 if (dbs_data->cdata->governor == GOV_ONDEMAND)
  68                         io_busy = od_tuners->io_is_busy;
  69                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time, io_busy);
  70
  71                 wall_time = (unsigned int)
  72                         (cur_wall_time - j_cdbs->prev_cpu_wall);
  73                 j_cdbs->prev_cpu_wall = cur_wall_time;
  74
  75                 idle_time = (unsigned int)
  76                         (cur_idle_time - j_cdbs->prev_cpu_idle);
  77                 j_cdbs->prev_cpu_idle = cur_idle_time;
  78
  79                 if (ignore_nice) {
  80                         u64 cur_nice;
  81                         unsigned long cur_nice_jiffies;
  82
  83                         cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] -
  84                                          cdbs->prev_cpu_nice;
  85                         /*
  86                          * Assumption: nice time between sampling periods will
  87                          * be less than 2^32 jiffies for 32 bit sys
  88                          */
  89                         cur_nice_jiffies = (unsigned long)
  90                                         cputime64_to_jiffies64(cur_nice);
  91
  92                         cdbs->prev_cpu_nice =
  93                                 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
  94                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
  95                 }
  96
  97                 if (unlikely(!wall_time || wall_time < idle_time))
  98                         continue;
  99
 100                 load = 100 * (wall_time - idle_time) / wall_time;
 101
 102                 if (load > max_load)
 103                         max_load = load;
 104         }
 105
 106         dbs_data->cdata->gov_check_cpu(cpu, max_load);
 107 }
 108 EXPORT_SYMBOL_GPL(dbs_check_cpu);
 109
 110 static inline void __gov_queue_work(int cpu, struct dbs_data *dbs_data,
 111                 unsigned int delay)
 112 {
 113         struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
 114
 115         mod_delayed_work_on(cpu, system_wq, &cdbs->work, delay);
 116 }
 117
 118 void gov_queue_work(struct dbs_data *dbs_data, struct cpufreq_policy *policy,
 119                 unsigned int delay, bool all_cpus)
 120 {
 121         int i;
 122
 123         if (!policy->governor_enabled)
 124                 return;
 125
 126         if (!all_cpus) {
 127                 __gov_queue_work(smp_processor_id(), dbs_data, delay);
 128         } else {
 129                 for_each_cpu(i, policy->cpus)
 130                         __gov_queue_work(i, dbs_data, delay);
 131         }
 132 }
 133 EXPORT_SYMBOL_GPL(gov_queue_work);
 134
 135 static inline void gov_cancel_work(struct dbs_data *dbs_data,
 136                 struct cpufreq_policy *policy)
 137 {
 138         struct cpu_dbs_common_info *cdbs;
 139         int i;
 140
 141         for_each_cpu(i, policy->cpus) {
 142                 cdbs = dbs_data->cdata->get_cpu_cdbs(i);
 143                 cancel_delayed_work_sync(&cdbs->work);
 144         }
 145 }
 146
 147 /* Will return if we need to evaluate cpu load again or not */
 148 bool need_load_eval(struct cpu_dbs_common_info *cdbs,
 149                 unsigned int sampling_rate)
 150 {
 151         if (policy_is_shared(cdbs->cur_policy)) {
 152                 ktime_t time_now = ktime_get();
 153                 s64 delta_us = ktime_us_delta(time_now, cdbs->time_stamp);
 154
 155                 /* Do nothing if we recently have sampled */
 156                 if (delta_us < (s64)(sampling_rate / 2))
 157                         return false;
 158                 else
 159                         cdbs->time_stamp = time_now;
 160         }
 161
 162         return true;
 163 }
 164 EXPORT_SYMBOL_GPL(need_load_eval);
 165
 166 static void set_sampling_rate(struct dbs_data *dbs_data,
 167                 unsigned int sampling_rate)
 168 {
 169         if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
 170                 struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
 171                 cs_tuners->sampling_rate = sampling_rate;
 172         } else {
 173                 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 174                 od_tuners->sampling_rate = sampling_rate;
 175         }
 176 }
 177
 178 int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 179                 struct common_dbs_data *cdata, unsigned int event)
 180 {
 181         struct dbs_data *dbs_data;
 182         struct od_cpu_dbs_info_s *od_dbs_info = NULL;
 183         struct cs_cpu_dbs_info_s *cs_dbs_info = NULL;
 184         struct od_ops *od_ops = NULL;
 185         struct od_dbs_tuners *od_tuners = NULL;
 186         struct cs_dbs_tuners *cs_tuners = NULL;
 187         struct cpu_dbs_common_info *cpu_cdbs;
 188         unsigned int sampling_rate, latency, ignore_nice, j, cpu = policy->cpu;
 189         int io_busy = 0;
 190         int rc;
 191
 192         if (have_governor_per_policy())
 193                 dbs_data = policy->governor_data;
 194         else
 195                 dbs_data = cdata->gdbs_data;
 196
 197         WARN_ON(!dbs_data && (event != CPUFREQ_GOV_POLICY_INIT));
 198
 199         switch (event) {
 200         case CPUFREQ_GOV_POLICY_INIT:
 201                 if (have_governor_per_policy()) {
 202                         WARN_ON(dbs_data);
 203                 } else if (dbs_data) {
 204                         dbs_data->usage_count++;
 205                         policy->governor_data = dbs_data;
 206                         return 0;
 207                 }
 208
 209                 dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
 210                 if (!dbs_data) {
 211                         pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
 212                         return -ENOMEM;
 213                 }
 214
 215                 dbs_data->cdata = cdata;
 216                 dbs_data->usage_count = 1;
 217                 rc = cdata->init(dbs_data);
 218                 if (rc) {
 219                         pr_err("%s: POLICY_INIT: init() failed\n", __func__);
 220                         kfree(dbs_data);
 221                         return rc;
 222                 }
 223
 224                 rc = sysfs_create_group(get_governor_parent_kobj(policy),
 225                                 get_sysfs_attr(dbs_data));
 226                 if (rc) {
 227                         cdata->exit(dbs_data);
 228                         kfree(dbs_data);
 229                         return rc;
 230                 }
 231
 232                 policy->governor_data = dbs_data;
 233
 234                 /* policy latency is in nS. Convert it to uS first */
 235                 latency = policy->cpuinfo.transition_latency / 1000;
 236                 if (latency == 0)
 237                         latency = 1;
 238
 239                 /* Bring kernel and HW constraints together */
 240                 dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate,
 241                                 MIN_LATENCY_MULTIPLIER * latency);
 242                 set_sampling_rate(dbs_data, max(dbs_data->min_sampling_rate,
 243                                         latency * LATENCY_MULTIPLIER));
 244
 245                 if ((cdata->governor == GOV_CONSERVATIVE) &&
 246                                 (!policy->governor->initialized)) {
 247                         struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
 248
 249                         cpufreq_register_notifier(cs_ops->notifier_block,
 250                                         CPUFREQ_TRANSITION_NOTIFIER);
 251                 }
 252
 253                 if (!have_governor_per_policy())
 254                         cdata->gdbs_data = dbs_data;
 255
 256                 return 0;
 257         case CPUFREQ_GOV_POLICY_EXIT:
 258                 if (!--dbs_data->usage_count) {
 259                         sysfs_remove_group(get_governor_parent_kobj(policy),
 260                                         get_sysfs_attr(dbs_data));
 261
 262                         if ((dbs_data->cdata->governor == GOV_CONSERVATIVE) &&
 263                                 (policy->governor->initialized == 1)) {
 264                                 struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
 265
 266                                 cpufreq_unregister_notifier(cs_ops->notifier_block,
 267                                                 CPUFREQ_TRANSITION_NOTIFIER);
 268                         }
 269
 270                         cdata->exit(dbs_data);
 271                         kfree(dbs_data);
 272                         cdata->gdbs_data = NULL;
 273                 }
 274
 275                 policy->governor_data = NULL;
 276                 return 0;
 277         }
 278
 279         cpu_cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
 280
 281         if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
 282                 cs_tuners = dbs_data->tuners;
 283                 cs_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
 284                 sampling_rate = cs_tuners->sampling_rate;
 285                 ignore_nice = cs_tuners->ignore_nice_load;
 286         } else {
 287                 od_tuners = dbs_data->tuners;
 288                 od_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
 289                 sampling_rate = od_tuners->sampling_rate;
 290                 ignore_nice = od_tuners->ignore_nice_load;
 291                 od_ops = dbs_data->cdata->gov_ops;
 292                 io_busy = od_tuners->io_is_busy;
 293         }
 294
 295         switch (event) {
 296         case CPUFREQ_GOV_START:
 297                 if (!policy->cur)
 298                         return -EINVAL;
 299
 300                 mutex_lock(&dbs_data->mutex);
 301
 302                 for_each_cpu(j, policy->cpus) {
 303                         struct cpu_dbs_common_info *j_cdbs =
 304                                 dbs_data->cdata->get_cpu_cdbs(j);
 305
 306                         j_cdbs->cpu = j;
 307                         j_cdbs->cur_policy = policy;
 308                         j_cdbs->prev_cpu_idle = get_cpu_idle_time(j,
 309                                                &j_cdbs->prev_cpu_wall, io_busy);
 310                         if (ignore_nice)
 311                                 j_cdbs->prev_cpu_nice =
 312                                         kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 313
 314                         mutex_init(&j_cdbs->timer_mutex);
 315                         INIT_DEFERRABLE_WORK(&j_cdbs->work,
 316                                              dbs_data->cdata->gov_dbs_timer);
 317                 }
 318
 319                 /*
 320                  * conservative does not implement micro like ondemand
 321                  * governor, thus we are bound to jiffes/HZ
 322                  */
 323                 if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
 324                         cs_dbs_info->down_skip = 0;
 325                         cs_dbs_info->enable = 1;
 326                         cs_dbs_info->requested_freq = policy->cur;
 327                 } else {
 328                         od_dbs_info->rate_mult = 1;
 329                         od_dbs_info->sample_type = OD_NORMAL_SAMPLE;
 330                         od_ops->powersave_bias_init_cpu(cpu);
 331                 }
 332
 333                 mutex_unlock(&dbs_data->mutex);
 334
 335                 /* Initiate timer time stamp */
 336                 cpu_cdbs->time_stamp = ktime_get();
 337
 338                 gov_queue_work(dbs_data, policy,
 339                                 delay_for_sampling_rate(sampling_rate), true);
 340                 break;
 341
 342         case CPUFREQ_GOV_STOP:
 343                 if (dbs_data->cdata->governor == GOV_CONSERVATIVE)
 344                         cs_dbs_info->enable = 0;
 345
 346                 gov_cancel_work(dbs_data, policy);
 347
 348                 mutex_lock(&dbs_data->mutex);
 349                 mutex_destroy(&cpu_cdbs->timer_mutex);
 350                 cpu_cdbs->cur_policy = NULL;
 351
 352                 mutex_unlock(&dbs_data->mutex);
 353
 354                 break;
 355
 356         case CPUFREQ_GOV_LIMITS:
 357                 mutex_lock(&dbs_data->mutex);
 358                 if (!cpu_cdbs->cur_policy) {
 359                         mutex_unlock(&dbs_data->mutex);
 360                         break;
 361                 }
 362                 mutex_lock(&cpu_cdbs->timer_mutex);
 363                 if (policy->max < cpu_cdbs->cur_policy->cur)
 364                         __cpufreq_driver_target(cpu_cdbs->cur_policy,
 365                                         policy->max, CPUFREQ_RELATION_H);
 366                 else if (policy->min > cpu_cdbs->cur_policy->cur)
 367                         __cpufreq_driver_target(cpu_cdbs->cur_policy,
 368                                         policy->min, CPUFREQ_RELATION_L);
 369                 dbs_check_cpu(dbs_data, cpu);
 370                 mutex_unlock(&cpu_cdbs->timer_mutex);
 371                 mutex_unlock(&dbs_data->mutex);
 372                 break;
 373         }
 374         return 0;
 375 }
 376 EXPORT_SYMBOL_GPL(cpufreq_governor_dbs);