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