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 <linux/export.h>
  20 #include <linux/kernel_stat.h>
  21 #include <linux/slab.h>
  22
  23 #include "cpufreq_governor.h"
  24
  25 static struct attribute_group *get_sysfs_attr(struct dbs_data *dbs_data)
  26 {
  27         if (have_governor_per_policy())
  28                 return dbs_data->cdata->attr_group_gov_pol;
  29         else
  30                 return dbs_data->cdata->attr_group_gov_sys;
  31 }
  32
  33 void dbs_check_cpu(struct dbs_data *dbs_data, int cpu)
  34 {
  35         struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
  36         struct od_dbs_tuners *od_tuners = dbs_data->tuners;
  37         struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
  38         struct cpufreq_policy *policy;
  39         unsigned int max_load = 0;
  40         unsigned int ignore_nice;
  41         unsigned int j;
  42
  43         if (dbs_data->cdata->governor == GOV_ONDEMAND)
  44                 ignore_nice = od_tuners->ignore_nice_load;
  45         else
  46                 ignore_nice = cs_tuners->ignore_nice_load;
  47
  48         policy = cdbs->cur_policy;
  49
  50         /* Get Absolute Load */
  51         for_each_cpu(j, policy->cpus) {
  52                 struct cpu_dbs_common_info *j_cdbs;
  53                 u64 cur_wall_time, cur_idle_time;
  54                 unsigned int idle_time, wall_time;
  55                 unsigned int load;
  56                 int io_busy = 0;
  57
  58                 j_cdbs = dbs_data->cdata->get_cpu_cdbs(j);
  59
  60                 /*
  61                  * For the purpose of ondemand, waiting for disk IO is
  62                  * an indication that you're performance critical, and
  63                  * not that the system is actually idle. So do not add
  64                  * the iowait time to the cpu idle time.
  65                  */
  66                 if (dbs_data->cdata->governor == GOV_ONDEMAND)
  67                         io_busy = od_tuners->io_is_busy;
  68                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time, io_busy);
  69
  70                 wall_time = (unsigned int)
  71                         (cur_wall_time - j_cdbs->prev_cpu_wall);
  72                 j_cdbs->prev_cpu_wall = cur_wall_time;
  73
  74                 idle_time = (unsigned int)
  75                         (cur_idle_time - j_cdbs->prev_cpu_idle);
  76                 j_cdbs->prev_cpu_idle = cur_idle_time;
  77
  78                 if (ignore_nice) {
  79                         u64 cur_nice;
  80                         unsigned long cur_nice_jiffies;
  81
  82                         cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] -
  83                                          cdbs->prev_cpu_nice;
  84                         /*
  85                          * Assumption: nice time between sampling periods will
  86                          * be less than 2^32 jiffies for 32 bit sys
  87                          */
  88                         cur_nice_jiffies = (unsigned long)
  89                                         cputime64_to_jiffies64(cur_nice);
  90
  91                         cdbs->prev_cpu_nice =
  92                                 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
  93                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
  94                 }
  95
  96                 if (unlikely(!wall_time || wall_time < idle_time))
  97                         continue;
  98
  99                 load = 100 * (wall_time - idle_time) / wall_time;
 100
 101                 if (load > max_load)
 102                         max_load = load;
 103         }
 104
 105         dbs_data->cdata->gov_check_cpu(cpu, max_load);
 106 }
 107 EXPORT_SYMBOL_GPL(dbs_check_cpu);
 108
 109 static inline void __gov_queue_work(int cpu, struct dbs_data *dbs_data,
 110                 unsigned int delay)
 111 {
 112         struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
 113
 114         mod_delayed_work_on(cpu, system_wq, &cdbs->work, delay);
 115 }
 116
 117 void gov_queue_work(struct dbs_data *dbs_data, struct cpufreq_policy *policy,
 118                 unsigned int delay, bool all_cpus)
 119 {
 120         int i;
 121
 122         if (!policy->governor_enabled)
 123                 return;
 124
 125         if (!all_cpus) {
 126                 /*
 127                  * Use raw_smp_processor_id() to avoid preemptible warnings.
 128                  * We know that this is only called with all_cpus == false from
 129                  * works that have been queued with *_work_on() functions and
 130                  * those works are canceled during CPU_DOWN_PREPARE so they
 131                  * can't possibly run on any other CPU.
 132                  */
 133                 __gov_queue_work(raw_smp_processor_id(), dbs_data, delay);
 134         } else {
 135                 for_each_cpu(i, policy->cpus)
 136                         __gov_queue_work(i, dbs_data, delay);
 137         }
 138 }
 139 EXPORT_SYMBOL_GPL(gov_queue_work);
 140
 141 static inline void gov_cancel_work(struct dbs_data *dbs_data,
 142                 struct cpufreq_policy *policy)
 143 {
 144         struct cpu_dbs_common_info *cdbs;
 145         int i;
 146
 147         for_each_cpu(i, policy->cpus) {
 148                 cdbs = dbs_data->cdata->get_cpu_cdbs(i);
 149                 cancel_delayed_work_sync(&cdbs->work);
 150         }
 151 }
 152
 153 /* Will return if we need to evaluate cpu load again or not */
 154 bool need_load_eval(struct cpu_dbs_common_info *cdbs,
 155                 unsigned int sampling_rate)
 156 {
 157         if (policy_is_shared(cdbs->cur_policy)) {
 158                 ktime_t time_now = ktime_get();
 159                 s64 delta_us = ktime_us_delta(time_now, cdbs->time_stamp);
 160
 161                 /* Do nothing if we recently have sampled */
 162                 if (delta_us < (s64)(sampling_rate / 2))
 163                         return false;
 164                 else
 165                         cdbs->time_stamp = time_now;
 166         }
 167
 168         return true;
 169 }
 170 EXPORT_SYMBOL_GPL(need_load_eval);
 171
 172 static void set_sampling_rate(struct dbs_data *dbs_data,
 173                 unsigned int sampling_rate)
 174 {
 175         if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
 176                 struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
 177                 cs_tuners->sampling_rate = sampling_rate;
 178         } else {
 179                 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 180                 od_tuners->sampling_rate = sampling_rate;
 181         }
 182 }
 183
 184 int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 185                 struct common_dbs_data *cdata, unsigned int event)
 186 {
 187         struct dbs_data *dbs_data;
 188         struct od_cpu_dbs_info_s *od_dbs_info = NULL;
 189         struct cs_cpu_dbs_info_s *cs_dbs_info = NULL;
 190         struct od_ops *od_ops = NULL;
 191         struct od_dbs_tuners *od_tuners = NULL;
 192         struct cs_dbs_tuners *cs_tuners = NULL;
 193         struct cpu_dbs_common_info *cpu_cdbs;
 194         unsigned int sampling_rate, latency, ignore_nice, j, cpu = policy->cpu;
 195         int io_busy = 0;
 196         int rc;
 197
 198         if (have_governor_per_policy())
 199                 dbs_data = policy->governor_data;
 200         else
 201                 dbs_data = cdata->gdbs_data;
 202
 203         WARN_ON(!dbs_data && (event != CPUFREQ_GOV_POLICY_INIT));
 204
 205         switch (event) {
 206         case CPUFREQ_GOV_POLICY_INIT:
 207                 if (have_governor_per_policy()) {
 208                         WARN_ON(dbs_data);
 209                 } else if (dbs_data) {
 210                         dbs_data->usage_count++;
 211                         policy->governor_data = dbs_data;
 212                         return 0;
 213                 }
 214
 215                 dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
 216                 if (!dbs_data) {
 217                         pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
 218                         return -ENOMEM;
 219                 }
 220
 221                 dbs_data->cdata = cdata;
 222                 dbs_data->usage_count = 1;
 223                 rc = cdata->init(dbs_data);
 224                 if (rc) {
 225                         pr_err("%s: POLICY_INIT: init() failed\n", __func__);
 226                         kfree(dbs_data);
 227                         return rc;
 228                 }
 229
 230                 if (!have_governor_per_policy())
 231                         WARN_ON(cpufreq_get_global_kobject());
 232
 233                 rc = sysfs_create_group(get_governor_parent_kobj(policy),
 234                                 get_sysfs_attr(dbs_data));
 235                 if (rc) {
 236                         cdata->exit(dbs_data);
 237                         kfree(dbs_data);
 238                         return rc;
 239                 }
 240
 241                 policy->governor_data = dbs_data;
 242
 243                 /* policy latency is in ns. Convert it to us first */
 244                 latency = policy->cpuinfo.transition_latency / 1000;
 245                 if (latency == 0)
 246                         latency = 1;
 247
 248                 /* Bring kernel and HW constraints together */
 249                 dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate,
 250                                 MIN_LATENCY_MULTIPLIER * latency);
 251                 set_sampling_rate(dbs_data, max(dbs_data->min_sampling_rate,
 252                                         latency * LATENCY_MULTIPLIER));
 253
 254                 if ((cdata->governor == GOV_CONSERVATIVE) &&
 255                                 (!policy->governor->initialized)) {
 256                         struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
 257
 258                         cpufreq_register_notifier(cs_ops->notifier_block,
 259                                         CPUFREQ_TRANSITION_NOTIFIER);
 260                 }
 261
 262                 if (!have_governor_per_policy())
 263                         cdata->gdbs_data = dbs_data;
 264
 265                 return 0;
 266         case CPUFREQ_GOV_POLICY_EXIT:
 267                 if (!--dbs_data->usage_count) {
 268                         sysfs_remove_group(get_governor_parent_kobj(policy),
 269                                         get_sysfs_attr(dbs_data));
 270
 271                         if (!have_governor_per_policy())
 272                                 cpufreq_put_global_kobject();
 273
 274                         if ((dbs_data->cdata->governor == GOV_CONSERVATIVE) &&
 275                                 (policy->governor->initialized == 1)) {
 276                                 struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
 277
 278                                 cpufreq_unregister_notifier(cs_ops->notifier_block,
 279                                                 CPUFREQ_TRANSITION_NOTIFIER);
 280                         }
 281
 282                         cdata->exit(dbs_data);
 283                         kfree(dbs_data);
 284                         cdata->gdbs_data = NULL;
 285                 }
 286
 287                 policy->governor_data = NULL;
 288                 return 0;
 289         }
 290
 291         cpu_cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
 292
 293         if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
 294                 cs_tuners = dbs_data->tuners;
 295                 cs_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
 296                 sampling_rate = cs_tuners->sampling_rate;
 297                 ignore_nice = cs_tuners->ignore_nice_load;
 298         } else {
 299                 od_tuners = dbs_data->tuners;
 300                 od_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
 301                 sampling_rate = od_tuners->sampling_rate;
 302                 ignore_nice = od_tuners->ignore_nice_load;
 303                 od_ops = dbs_data->cdata->gov_ops;
 304                 io_busy = od_tuners->io_is_busy;
 305         }
 306
 307         switch (event) {
 308         case CPUFREQ_GOV_START:
 309                 if (!policy->cur)
 310                         return -EINVAL;
 311
 312                 mutex_lock(&dbs_data->mutex);
 313
 314                 for_each_cpu(j, policy->cpus) {
 315                         struct cpu_dbs_common_info *j_cdbs =
 316                                 dbs_data->cdata->get_cpu_cdbs(j);
 317
 318                         j_cdbs->cpu = j;
 319                         j_cdbs->cur_policy = policy;
 320                         j_cdbs->prev_cpu_idle = get_cpu_idle_time(j,
 321                                                &j_cdbs->prev_cpu_wall, io_busy);
 322                         if (ignore_nice)
 323                                 j_cdbs->prev_cpu_nice =
 324                                         kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 325
 326                         mutex_init(&j_cdbs->timer_mutex);
 327                         INIT_DEFERRABLE_WORK(&j_cdbs->work,
 328                                              dbs_data->cdata->gov_dbs_timer);
 329                 }
 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);