kernel/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 sampling_rate;
  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                 struct od_cpu_dbs_info_s *od_dbs_info =
  46                                 dbs_data->cdata->get_cpu_dbs_info_s(cpu);
  47
  48                 /*
  49                  * Sometimes, the ondemand governor uses an additional
  50                  * multiplier to give long delays. So apply this multiplier to
  51                  * the 'sampling_rate', so as to keep the wake-up-from-idle
  52                  * detection logic a bit conservative.
  53                  */
  54                 sampling_rate = od_tuners->sampling_rate;
  55                 sampling_rate *= od_dbs_info->rate_mult;
  56
  57                 ignore_nice = od_tuners->ignore_nice_load;
  58         } else {
  59                 sampling_rate = cs_tuners->sampling_rate;
  60                 ignore_nice = cs_tuners->ignore_nice_load;
  61         }
  62
  63         policy = cdbs->cur_policy;
  64
  65         /* Get Absolute Load */
  66         for_each_cpu(j, policy->cpus) {
  67                 struct cpu_dbs_common_info *j_cdbs;
  68                 u64 cur_wall_time, cur_idle_time;
  69                 unsigned int idle_time, wall_time;
  70                 unsigned int load;
  71                 int io_busy = 0;
  72
  73                 j_cdbs = dbs_data->cdata->get_cpu_cdbs(j);
  74
  75                 /*
  76                  * For the purpose of ondemand, waiting for disk IO is
  77                  * an indication that you're performance critical, and
  78                  * not that the system is actually idle. So do not add
  79                  * the iowait time to the cpu idle time.
  80                  */
  81                 if (dbs_data->cdata->governor == GOV_ONDEMAND)
  82                         io_busy = od_tuners->io_is_busy;
  83                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time, io_busy);
  84
  85                 wall_time = (unsigned int)
  86                         (cur_wall_time - j_cdbs->prev_cpu_wall);
  87                 j_cdbs->prev_cpu_wall = cur_wall_time;
  88
  89                 idle_time = (unsigned int)
  90                         (cur_idle_time - j_cdbs->prev_cpu_idle);
  91                 j_cdbs->prev_cpu_idle = cur_idle_time;
  92
  93                 if (ignore_nice) {
  94                         u64 cur_nice;
  95                         unsigned long cur_nice_jiffies;
  96
  97                         cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] -
  98                                          cdbs->prev_cpu_nice;
  99                         /*
 100                          * Assumption: nice time between sampling periods will
 101                          * be less than 2^32 jiffies for 32 bit sys
 102                          */
 103                         cur_nice_jiffies = (unsigned long)
 104                                         cputime64_to_jiffies64(cur_nice);
 105
 106                         cdbs->prev_cpu_nice =
 107                                 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 108                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
 109                 }
 110
 111                 if (unlikely(!wall_time || wall_time < idle_time))
 112                         continue;
 113
 114                 /*
 115                  * If the CPU had gone completely idle, and a task just woke up
 116                  * on this CPU now, it would be unfair to calculate 'load' the
 117                  * usual way for this elapsed time-window, because it will show
 118                  * near-zero load, irrespective of how CPU intensive that task
 119                  * actually is. This is undesirable for latency-sensitive bursty
 120                  * workloads.
 121                  *
 122                  * To avoid this, we reuse the 'load' from the previous
 123                  * time-window and give this task a chance to start with a
 124                  * reasonably high CPU frequency. (However, we shouldn't over-do
 125                  * this copy, lest we get stuck at a high load (high frequency)
 126                  * for too long, even when the current system load has actually
 127                  * dropped down. So we perform the copy only once, upon the
 128                  * first wake-up from idle.)
 129                  *
 130                  * Detecting this situation is easy: the governor's deferrable
 131                  * timer would not have fired during CPU-idle periods. Hence
 132                  * an unusually large 'wall_time' (as compared to the sampling
 133                  * rate) indicates this scenario.
 134                  *
 135                  * prev_load can be zero in two cases and we must recalculate it
 136                  * for both cases:
 137                  * - during long idle intervals
 138                  * - explicitly set to zero
 139                  */
 140                 if (unlikely(wall_time > (2 * sampling_rate) &&
 141                              j_cdbs->prev_load)) {
 142                         load = j_cdbs->prev_load;
 143
 144                         /*
 145                          * Perform a destructive copy, to ensure that we copy
 146                          * the previous load only once, upon the first wake-up
 147                          * from idle.
 148                          */
 149                         j_cdbs->prev_load = 0;
 150                 } else {
 151                         load = 100 * (wall_time - idle_time) / wall_time;
 152                         j_cdbs->prev_load = load;
 153                 }
 154
 155                 if (load > max_load)
 156                         max_load = load;
 157         }
 158
 159         dbs_data->cdata->gov_check_cpu(cpu, max_load);
 160 }
 161 EXPORT_SYMBOL_GPL(dbs_check_cpu);
 162
 163 static inline void __gov_queue_work(int cpu, struct dbs_data *dbs_data,
 164                 unsigned int delay)
 165 {
 166         struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
 167
 168         mod_delayed_work_on(cpu, system_wq, &cdbs->work, delay);
 169 }
 170
 171 void gov_queue_work(struct dbs_data *dbs_data, struct cpufreq_policy *policy,
 172                 unsigned int delay, bool all_cpus)
 173 {
 174         int i;
 175
 176         mutex_lock(&cpufreq_governor_lock);
 177         if (!policy->governor_enabled)
 178                 goto out_unlock;
 179
 180         if (!all_cpus) {
 181                 /*
 182                  * Use raw_smp_processor_id() to avoid preemptible warnings.
 183                  * We know that this is only called with all_cpus == false from
 184                  * works that have been queued with *_work_on() functions and
 185                  * those works are canceled during CPU_DOWN_PREPARE so they
 186                  * can't possibly run on any other CPU.
 187                  */
 188                 __gov_queue_work(raw_smp_processor_id(), dbs_data, delay);
 189         } else {
 190                 for_each_cpu(i, policy->cpus)
 191                         __gov_queue_work(i, dbs_data, delay);
 192         }
 193
 194 out_unlock:
 195         mutex_unlock(&cpufreq_governor_lock);
 196 }
 197 EXPORT_SYMBOL_GPL(gov_queue_work);
 198
 199 static inline void gov_cancel_work(struct dbs_data *dbs_data,
 200                 struct cpufreq_policy *policy)
 201 {
 202         struct cpu_dbs_common_info *cdbs;
 203         int i;
 204
 205         for_each_cpu(i, policy->cpus) {
 206                 cdbs = dbs_data->cdata->get_cpu_cdbs(i);
 207                 cancel_delayed_work_sync(&cdbs->work);
 208         }
 209 }
 210
 211 /* Will return if we need to evaluate cpu load again or not */
 212 bool need_load_eval(struct cpu_dbs_common_info *cdbs,
 213                 unsigned int sampling_rate)
 214 {
 215         if (policy_is_shared(cdbs->cur_policy)) {
 216                 ktime_t time_now = ktime_get();
 217                 s64 delta_us = ktime_us_delta(time_now, cdbs->time_stamp);
 218
 219                 /* Do nothing if we recently have sampled */
 220                 if (delta_us < (s64)(sampling_rate / 2))
 221                         return false;
 222                 else
 223                         cdbs->time_stamp = time_now;
 224         }
 225
 226         return true;
 227 }
 228 EXPORT_SYMBOL_GPL(need_load_eval);
 229
 230 static void set_sampling_rate(struct dbs_data *dbs_data,
 231                 unsigned int sampling_rate)
 232 {
 233         if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
 234                 struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
 235                 cs_tuners->sampling_rate = sampling_rate;
 236         } else {
 237                 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 238                 od_tuners->sampling_rate = sampling_rate;
 239         }
 240 }
 241
 242 int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 243                 struct common_dbs_data *cdata, unsigned int event)
 244 {
 245         struct dbs_data *dbs_data;
 246         struct od_cpu_dbs_info_s *od_dbs_info = NULL;
 247         struct cs_cpu_dbs_info_s *cs_dbs_info = NULL;
 248         struct od_ops *od_ops = NULL;
 249         struct od_dbs_tuners *od_tuners = NULL;
 250         struct cs_dbs_tuners *cs_tuners = NULL;
 251         struct cpu_dbs_common_info *cpu_cdbs;
 252         unsigned int sampling_rate, latency, ignore_nice, j, cpu = policy->cpu;
 253         int io_busy = 0;
 254         int rc;
 255
 256         if (have_governor_per_policy())
 257                 dbs_data = policy->governor_data;
 258         else
 259                 dbs_data = cdata->gdbs_data;
 260
 261         WARN_ON(!dbs_data && (event != CPUFREQ_GOV_POLICY_INIT));
 262
 263         switch (event) {
 264         case CPUFREQ_GOV_POLICY_INIT:
 265                 if (have_governor_per_policy()) {
 266                         WARN_ON(dbs_data);
 267                 } else if (dbs_data) {
 268                         dbs_data->usage_count++;
 269                         policy->governor_data = dbs_data;
 270                         return 0;
 271                 }
 272
 273                 dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
 274                 if (!dbs_data) {
 275                         pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
 276                         return -ENOMEM;
 277                 }
 278
 279                 dbs_data->cdata = cdata;
 280                 dbs_data->usage_count = 1;
 281                 rc = cdata->init(dbs_data);
 282                 if (rc) {
 283                         pr_err("%s: POLICY_INIT: init() failed\n", __func__);
 284                         kfree(dbs_data);
 285                         return rc;
 286                 }
 287
 288                 if (!have_governor_per_policy())
 289                         WARN_ON(cpufreq_get_global_kobject());
 290
 291                 rc = sysfs_create_group(get_governor_parent_kobj(policy),
 292                                 get_sysfs_attr(dbs_data));
 293                 if (rc) {
 294                         cdata->exit(dbs_data);
 295                         kfree(dbs_data);
 296                         return rc;
 297                 }
 298
 299                 policy->governor_data = dbs_data;
 300
 301                 /* policy latency is in ns. Convert it to us first */
 302                 latency = policy->cpuinfo.transition_latency / 1000;
 303                 if (latency == 0)
 304                         latency = 1;
 305
 306                 /* Bring kernel and HW constraints together */
 307                 dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate,
 308                                 MIN_LATENCY_MULTIPLIER * latency);
 309                 set_sampling_rate(dbs_data, max(dbs_data->min_sampling_rate,
 310                                         latency * LATENCY_MULTIPLIER));
 311
 312                 if ((cdata->governor == GOV_CONSERVATIVE) &&
 313                                 (!policy->governor->initialized)) {
 314                         struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
 315
 316                         cpufreq_register_notifier(cs_ops->notifier_block,
 317                                         CPUFREQ_TRANSITION_NOTIFIER);
 318                 }
 319
 320                 if (!have_governor_per_policy())
 321                         cdata->gdbs_data = dbs_data;
 322
 323                 return 0;
 324         case CPUFREQ_GOV_POLICY_EXIT:
 325                 if (!--dbs_data->usage_count) {
 326                         sysfs_remove_group(get_governor_parent_kobj(policy),
 327                                         get_sysfs_attr(dbs_data));
 328
 329                         if (!have_governor_per_policy())
 330                                 cpufreq_put_global_kobject();
 331
 332                         if ((dbs_data->cdata->governor == GOV_CONSERVATIVE) &&
 333                                 (policy->governor->initialized == 1)) {
 334                                 struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
 335
 336                                 cpufreq_unregister_notifier(cs_ops->notifier_block,
 337                                                 CPUFREQ_TRANSITION_NOTIFIER);
 338                         }
 339
 340                         cdata->exit(dbs_data);
 341                         kfree(dbs_data);
 342                         cdata->gdbs_data = NULL;
 343                 }
 344
 345                 policy->governor_data = NULL;
 346                 return 0;
 347         }
 348
 349         cpu_cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
 350
 351         if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
 352                 cs_tuners = dbs_data->tuners;
 353                 cs_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
 354                 sampling_rate = cs_tuners->sampling_rate;
 355                 ignore_nice = cs_tuners->ignore_nice_load;
 356         } else {
 357                 od_tuners = dbs_data->tuners;
 358                 od_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
 359                 sampling_rate = od_tuners->sampling_rate;
 360                 ignore_nice = od_tuners->ignore_nice_load;
 361                 od_ops = dbs_data->cdata->gov_ops;
 362                 io_busy = od_tuners->io_is_busy;
 363         }
 364
 365         switch (event) {
 366         case CPUFREQ_GOV_START:
 367                 if (!policy->cur)
 368                         return -EINVAL;
 369
 370                 mutex_lock(&dbs_data->mutex);
 371
 372                 for_each_cpu(j, policy->cpus) {
 373                         struct cpu_dbs_common_info *j_cdbs =
 374                                 dbs_data->cdata->get_cpu_cdbs(j);
 375                         unsigned int prev_load;
 376
 377                         j_cdbs->cpu = j;
 378                         j_cdbs->cur_policy = policy;
 379                         j_cdbs->prev_cpu_idle = get_cpu_idle_time(j,
 380                                                &j_cdbs->prev_cpu_wall, io_busy);
 381
 382                         prev_load = (unsigned int)
 383                                 (j_cdbs->prev_cpu_wall - j_cdbs->prev_cpu_idle);
 384                         j_cdbs->prev_load = 100 * prev_load /
 385                                         (unsigned int) j_cdbs->prev_cpu_wall;
 386
 387                         if (ignore_nice)
 388                                 j_cdbs->prev_cpu_nice =
 389                                         kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 390
 391                         mutex_init(&j_cdbs->timer_mutex);
 392                         INIT_DEFERRABLE_WORK(&j_cdbs->work,
 393                                              dbs_data->cdata->gov_dbs_timer);
 394                 }
 395
 396                 if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
 397                         cs_dbs_info->down_skip = 0;
 398                         cs_dbs_info->enable = 1;
 399                         cs_dbs_info->requested_freq = policy->cur;
 400                 } else {
 401                         od_dbs_info->rate_mult = 1;
 402                         od_dbs_info->sample_type = OD_NORMAL_SAMPLE;
 403                         od_ops->powersave_bias_init_cpu(cpu);
 404                 }
 405
 406                 mutex_unlock(&dbs_data->mutex);
 407
 408                 /* Initiate timer time stamp */
 409                 cpu_cdbs->time_stamp = ktime_get();
 410
 411                 gov_queue_work(dbs_data, policy,
 412                                 delay_for_sampling_rate(sampling_rate), true);
 413                 break;
 414
 415         case CPUFREQ_GOV_STOP:
 416                 if (dbs_data->cdata->governor == GOV_CONSERVATIVE)
 417                         cs_dbs_info->enable = 0;
 418
 419                 gov_cancel_work(dbs_data, policy);
 420
 421                 mutex_lock(&dbs_data->mutex);
 422                 mutex_destroy(&cpu_cdbs->timer_mutex);
 423                 cpu_cdbs->cur_policy = NULL;
 424
 425                 mutex_unlock(&dbs_data->mutex);
 426
 427                 break;
 428
 429         case CPUFREQ_GOV_LIMITS:
 430                 mutex_lock(&dbs_data->mutex);
 431                 if (!cpu_cdbs->cur_policy) {
 432                         mutex_unlock(&dbs_data->mutex);
 433                         break;
 434                 }
 435                 mutex_lock(&cpu_cdbs->timer_mutex);
 436                 if (policy->max < cpu_cdbs->cur_policy->cur)
 437                         __cpufreq_driver_target(cpu_cdbs->cur_policy,
 438                                         policy->max, CPUFREQ_RELATION_H);
 439                 else if (policy->min > cpu_cdbs->cur_policy->cur)
 440                         __cpufreq_driver_target(cpu_cdbs->cur_policy,
 441                                         policy->min, CPUFREQ_RELATION_L);
 442                 dbs_check_cpu(dbs_data, cpu);
 443                 mutex_unlock(&cpu_cdbs->timer_mutex);
 444                 mutex_unlock(&dbs_data->mutex);
 445                 break;
 446         }
 447         return 0;
 448 }
 449 EXPORT_SYMBOL_GPL(cpufreq_governor_dbs);