drivers/cpufreq/cpufreq_ondemand.c

   1 /*
   2  *  drivers/cpufreq/cpufreq_ondemand.c
   3  *
   4  *  Copyright (C)  2001 Russell King
   5  *            (C)  2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
   6  *                      Jun Nakajima <jun.nakajima@intel.com>
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 as
  10  * published by the Free Software Foundation.
  11  */
  12
  13 #include <linux/kernel.h>
  14 #include <linux/module.h>
  15 #include <linux/init.h>
  16 #include <linux/cpufreq.h>
  17 #include <linux/cpu.h>
  18 #include <linux/jiffies.h>
  19 #include <linux/kernel_stat.h>
  20 #include <linux/mutex.h>
  21 #include <linux/hrtimer.h>
  22 #include <linux/tick.h>
  23 #include <linux/ktime.h>
  24
  25 /*
  26  * dbs is used in this file as a shortform for demandbased switching
  27  * It helps to keep variable names smaller, simpler
  28  */
  29
  30 #define DEF_FREQUENCY_DOWN_DIFFERENTIAL         (10)
  31 #define DEF_FREQUENCY_UP_THRESHOLD              (80)
  32 #define MICRO_FREQUENCY_DOWN_DIFFERENTIAL       (3)
  33 #define MICRO_FREQUENCY_UP_THRESHOLD            (95)
  34 #define MIN_FREQUENCY_UP_THRESHOLD              (11)
  35 #define MAX_FREQUENCY_UP_THRESHOLD              (100)
  36
  37 /*
  38  * The polling frequency of this governor depends on the capability of
  39  * the processor. Default polling frequency is 1000 times the transition
  40  * latency of the processor. The governor will work on any processor with
  41  * transition latency <= 10mS, using appropriate sampling
  42  * rate.
  43  * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
  44  * this governor will not work.
  45  * All times here are in uS.
  46  */
  47 static unsigned int def_sampling_rate;
  48 #define MIN_SAMPLING_RATE_RATIO                 (2)
  49 /* for correct statistics, we need at least 10 ticks between each measure */
  50 #define MIN_STAT_SAMPLING_RATE                  \
  51                         (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
  52 #define MIN_SAMPLING_RATE                       \
  53                         (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
  54 #define MAX_SAMPLING_RATE                       (500 * def_sampling_rate)
  55 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER    (1000)
  56 #define TRANSITION_LATENCY_LIMIT                (10 * 1000 * 1000)
  57
  58 static void do_dbs_timer(struct work_struct *work);
  59
  60 /* Sampling types */
  61 enum {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE};
  62
  63 struct cpu_dbs_info_s {
  64         cputime64_t prev_cpu_idle;
  65         cputime64_t prev_cpu_wall;
  66         cputime64_t prev_cpu_nice;
  67         struct cpufreq_policy *cur_policy;
  68         struct delayed_work work;
  69         struct cpufreq_frequency_table *freq_table;
  70         unsigned int freq_lo;
  71         unsigned int freq_lo_jiffies;
  72         unsigned int freq_hi_jiffies;
  73         int cpu;
  74         unsigned int enable:1,
  75                      sample_type:1;
  76 };
  77 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
  78
  79 static unsigned int dbs_enable; /* number of CPUs using this policy */
  80
  81 /*
  82  * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug
  83  * lock and dbs_mutex. cpu_hotplug lock should always be held before
  84  * dbs_mutex. If any function that can potentially take cpu_hotplug lock
  85  * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
  86  * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
  87  * is recursive for the same process. -Venki
  88  */
  89 static DEFINE_MUTEX(dbs_mutex);
  90
  91 static struct workqueue_struct  *kondemand_wq;
  92
  93 static struct dbs_tuners {
  94         unsigned int sampling_rate;
  95         unsigned int up_threshold;
  96         unsigned int down_differential;
  97         unsigned int ignore_nice;
  98         unsigned int powersave_bias;
  99 } dbs_tuners_ins = {
 100         .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
 101         .down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL,
 102         .ignore_nice = 0,
 103         .powersave_bias = 0,
 104 };
 105
 106 static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
 107                                                         cputime64_t *wall)
 108 {
 109         cputime64_t idle_time;
 110         cputime64_t cur_wall_time;
 111         cputime64_t busy_time;
 112
 113         cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
 114         busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
 115                         kstat_cpu(cpu).cpustat.system);
 116
 117         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
 118         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
 119         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
 120         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice);
 121
 122         idle_time = cputime64_sub(cur_wall_time, busy_time);
 123         if (wall)
 124                 *wall = cur_wall_time;
 125
 126         return idle_time;
 127 }
 128
 129 static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
 130 {
 131         u64 idle_time = get_cpu_idle_time_us(cpu, wall);
 132
 133         if (idle_time == -1ULL)
 134                 return get_cpu_idle_time_jiffy(cpu, wall);
 135
 136         return idle_time;
 137 }
 138
 139 /*
 140  * Find right freq to be set now with powersave_bias on.
 141  * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
 142  * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
 143  */
 144 static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
 145                                           unsigned int freq_next,
 146                                           unsigned int relation)
 147 {
 148         unsigned int freq_req, freq_reduc, freq_avg;
 149         unsigned int freq_hi, freq_lo;
 150         unsigned int index = 0;
 151         unsigned int jiffies_total, jiffies_hi, jiffies_lo;
 152         struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, policy->cpu);
 153
 154         if (!dbs_info->freq_table) {
 155                 dbs_info->freq_lo = 0;
 156                 dbs_info->freq_lo_jiffies = 0;
 157                 return freq_next;
 158         }
 159
 160         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
 161                         relation, &index);
 162         freq_req = dbs_info->freq_table[index].frequency;
 163         freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
 164         freq_avg = freq_req - freq_reduc;
 165
 166         /* Find freq bounds for freq_avg in freq_table */
 167         index = 0;
 168         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 169                         CPUFREQ_RELATION_H, &index);
 170         freq_lo = dbs_info->freq_table[index].frequency;
 171         index = 0;
 172         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 173                         CPUFREQ_RELATION_L, &index);
 174         freq_hi = dbs_info->freq_table[index].frequency;
 175
 176         /* Find out how long we have to be in hi and lo freqs */
 177         if (freq_hi == freq_lo) {
 178                 dbs_info->freq_lo = 0;
 179                 dbs_info->freq_lo_jiffies = 0;
 180                 return freq_lo;
 181         }
 182         jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 183         jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
 184         jiffies_hi += ((freq_hi - freq_lo) / 2);
 185         jiffies_hi /= (freq_hi - freq_lo);
 186         jiffies_lo = jiffies_total - jiffies_hi;
 187         dbs_info->freq_lo = freq_lo;
 188         dbs_info->freq_lo_jiffies = jiffies_lo;
 189         dbs_info->freq_hi_jiffies = jiffies_hi;
 190         return freq_hi;
 191 }
 192
 193 static void ondemand_powersave_bias_init(void)
 194 {
 195         int i;
 196         for_each_online_cpu(i) {
 197                 struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, i);
 198                 dbs_info->freq_table = cpufreq_frequency_get_table(i);
 199                 dbs_info->freq_lo = 0;
 200         }
 201 }
 202
 203 /************************** sysfs interface ************************/
 204 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
 205 {
 206         return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
 207 }
 208
 209 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
 210 {
 211         return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
 212 }
 213
 214 #define define_one_ro(_name)            \
 215 static struct freq_attr _name =         \
 216 __ATTR(_name, 0444, show_##_name, NULL)
 217
 218 define_one_ro(sampling_rate_max);
 219 define_one_ro(sampling_rate_min);
 220
 221 /* cpufreq_ondemand Governor Tunables */
 222 #define show_one(file_name, object)                                     \
 223 static ssize_t show_##file_name                                         \
 224 (struct cpufreq_policy *unused, char *buf)                              \
 225 {                                                                       \
 226         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
 227 }
 228 show_one(sampling_rate, sampling_rate);
 229 show_one(up_threshold, up_threshold);
 230 show_one(ignore_nice_load, ignore_nice);
 231 show_one(powersave_bias, powersave_bias);
 232
 233 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
 234                 const char *buf, size_t count)
 235 {
 236         unsigned int input;
 237         int ret;
 238         ret = sscanf(buf, "%u", &input);
 239
 240         mutex_lock(&dbs_mutex);
 241         if (ret != 1 || input > MAX_SAMPLING_RATE
 242                      || input < MIN_SAMPLING_RATE) {
 243                 mutex_unlock(&dbs_mutex);
 244                 return -EINVAL;
 245         }
 246
 247         dbs_tuners_ins.sampling_rate = input;
 248         mutex_unlock(&dbs_mutex);
 249
 250         return count;
 251 }
 252
 253 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
 254                 const char *buf, size_t count)
 255 {
 256         unsigned int input;
 257         int ret;
 258         ret = sscanf(buf, "%u", &input);
 259
 260         mutex_lock(&dbs_mutex);
 261         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 262                         input < MIN_FREQUENCY_UP_THRESHOLD) {
 263                 mutex_unlock(&dbs_mutex);
 264                 return -EINVAL;
 265         }
 266
 267         dbs_tuners_ins.up_threshold = input;
 268         mutex_unlock(&dbs_mutex);
 269
 270         return count;
 271 }
 272
 273 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
 274                 const char *buf, size_t count)
 275 {
 276         unsigned int input;
 277         int ret;
 278
 279         unsigned int j;
 280
 281         ret = sscanf(buf, "%u", &input);
 282         if ( ret != 1 )
 283                 return -EINVAL;
 284
 285         if ( input > 1 )
 286                 input = 1;
 287
 288         mutex_lock(&dbs_mutex);
 289         if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
 290                 mutex_unlock(&dbs_mutex);
 291                 return count;
 292         }
 293         dbs_tuners_ins.ignore_nice = input;
 294
 295         /* we need to re-evaluate prev_cpu_idle */
 296         for_each_online_cpu(j) {
 297                 struct cpu_dbs_info_s *dbs_info;
 298                 dbs_info = &per_cpu(cpu_dbs_info, j);
 299                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 300                                                 &dbs_info->prev_cpu_wall);
 301                 if (dbs_tuners_ins.ignore_nice)
 302                         dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 303
 304         }
 305         mutex_unlock(&dbs_mutex);
 306
 307         return count;
 308 }
 309
 310 static ssize_t store_powersave_bias(struct cpufreq_policy *unused,
 311                 const char *buf, size_t count)
 312 {
 313         unsigned int input;
 314         int ret;
 315         ret = sscanf(buf, "%u", &input);
 316
 317         if (ret != 1)
 318                 return -EINVAL;
 319
 320         if (input > 1000)
 321                 input = 1000;
 322
 323         mutex_lock(&dbs_mutex);
 324         dbs_tuners_ins.powersave_bias = input;
 325         ondemand_powersave_bias_init();
 326         mutex_unlock(&dbs_mutex);
 327
 328         return count;
 329 }
 330
 331 #define define_one_rw(_name) \
 332 static struct freq_attr _name = \
 333 __ATTR(_name, 0644, show_##_name, store_##_name)
 334
 335 define_one_rw(sampling_rate);
 336 define_one_rw(up_threshold);
 337 define_one_rw(ignore_nice_load);
 338 define_one_rw(powersave_bias);
 339
 340 static struct attribute * dbs_attributes[] = {
 341         &sampling_rate_max.attr,
 342         &sampling_rate_min.attr,
 343         &sampling_rate.attr,
 344         &up_threshold.attr,
 345         &ignore_nice_load.attr,
 346         &powersave_bias.attr,
 347         NULL
 348 };
 349
 350 static struct attribute_group dbs_attr_group = {
 351         .attrs = dbs_attributes,
 352         .name = "ondemand",
 353 };
 354
 355 /************************** sysfs end ************************/
 356
 357 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
 358 {
 359         unsigned int max_load_freq;
 360
 361         struct cpufreq_policy *policy;
 362         unsigned int j;
 363
 364         if (!this_dbs_info->enable)
 365                 return;
 366
 367         this_dbs_info->freq_lo = 0;
 368         policy = this_dbs_info->cur_policy;
 369
 370         /*
 371          * Every sampling_rate, we check, if current idle time is less
 372          * than 20% (default), then we try to increase frequency
 373          * Every sampling_rate, we look for a the lowest
 374          * frequency which can sustain the load while keeping idle time over
 375          * 30%. If such a frequency exist, we try to decrease to this frequency.
 376          *
 377          * Any frequency increase takes it to the maximum frequency.
 378          * Frequency reduction happens at minimum steps of
 379          * 5% (default) of current frequency
 380          */
 381
 382         /* Get Absolute Load - in terms of freq */
 383         max_load_freq = 0;
 384
 385         for_each_cpu(j, policy->cpus) {
 386                 struct cpu_dbs_info_s *j_dbs_info;
 387                 cputime64_t cur_wall_time, cur_idle_time;
 388                 unsigned int idle_time, wall_time;
 389                 unsigned int load, load_freq;
 390                 int freq_avg;
 391
 392                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
 393
 394                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
 395
 396                 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
 397                                 j_dbs_info->prev_cpu_wall);
 398                 j_dbs_info->prev_cpu_wall = cur_wall_time;
 399
 400                 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
 401                                 j_dbs_info->prev_cpu_idle);
 402                 j_dbs_info->prev_cpu_idle = cur_idle_time;
 403
 404                 if (dbs_tuners_ins.ignore_nice) {
 405                         cputime64_t cur_nice;
 406                         unsigned long cur_nice_jiffies;
 407
 408                         cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
 409                                          j_dbs_info->prev_cpu_nice);
 410                         /*
 411                          * Assumption: nice time between sampling periods will
 412                          * be less than 2^32 jiffies for 32 bit sys
 413                          */
 414                         cur_nice_jiffies = (unsigned long)
 415                                         cputime64_to_jiffies64(cur_nice);
 416
 417                         j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 418                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
 419                 }
 420
 421                 if (unlikely(!wall_time || wall_time < idle_time))
 422                         continue;
 423
 424                 load = 100 * (wall_time - idle_time) / wall_time;
 425
 426                 freq_avg = __cpufreq_driver_getavg(policy, j);
 427                 if (freq_avg <= 0)
 428                         freq_avg = policy->cur;
 429
 430                 load_freq = load * freq_avg;
 431                 if (load_freq > max_load_freq)
 432                         max_load_freq = load_freq;
 433         }
 434
 435         /* Check for frequency increase */
 436         if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
 437                 /* if we are already at full speed then break out early */
 438                 if (!dbs_tuners_ins.powersave_bias) {
 439                         if (policy->cur == policy->max)
 440                                 return;
 441
 442                         __cpufreq_driver_target(policy, policy->max,
 443                                 CPUFREQ_RELATION_H);
 444                 } else {
 445                         int freq = powersave_bias_target(policy, policy->max,
 446                                         CPUFREQ_RELATION_H);
 447                         __cpufreq_driver_target(policy, freq,
 448                                 CPUFREQ_RELATION_L);
 449                 }
 450                 return;
 451         }
 452
 453         /* Check for frequency decrease */
 454         /* if we cannot reduce the frequency anymore, break out early */
 455         if (policy->cur == policy->min)
 456                 return;
 457
 458         /*
 459          * The optimal frequency is the frequency that is the lowest that
 460          * can support the current CPU usage without triggering the up
 461          * policy. To be safe, we focus 10 points under the threshold.
 462          */
 463         if (max_load_freq <
 464             (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
 465              policy->cur) {
 466                 unsigned int freq_next;
 467                 freq_next = max_load_freq /
 468                                 (dbs_tuners_ins.up_threshold -
 469                                  dbs_tuners_ins.down_differential);
 470
 471                 if (!dbs_tuners_ins.powersave_bias) {
 472                         __cpufreq_driver_target(policy, freq_next,
 473                                         CPUFREQ_RELATION_L);
 474                 } else {
 475                         int freq = powersave_bias_target(policy, freq_next,
 476                                         CPUFREQ_RELATION_L);
 477                         __cpufreq_driver_target(policy, freq,
 478                                 CPUFREQ_RELATION_L);
 479                 }
 480         }
 481 }
 482
 483 static void do_dbs_timer(struct work_struct *work)
 484 {
 485         struct cpu_dbs_info_s *dbs_info =
 486                 container_of(work, struct cpu_dbs_info_s, work.work);
 487         unsigned int cpu = dbs_info->cpu;
 488         int sample_type = dbs_info->sample_type;
 489
 490         /* We want all CPUs to do sampling nearly on same jiffy */
 491         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 492
 493         delay -= jiffies % delay;
 494
 495         if (lock_policy_rwsem_write(cpu) < 0)
 496                 return;
 497
 498         if (!dbs_info->enable) {
 499                 unlock_policy_rwsem_write(cpu);
 500                 return;
 501         }
 502
 503         /* Common NORMAL_SAMPLE setup */
 504         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 505         if (!dbs_tuners_ins.powersave_bias ||
 506             sample_type == DBS_NORMAL_SAMPLE) {
 507                 dbs_check_cpu(dbs_info);
 508                 if (dbs_info->freq_lo) {
 509                         /* Setup timer for SUB_SAMPLE */
 510                         dbs_info->sample_type = DBS_SUB_SAMPLE;
 511                         delay = dbs_info->freq_hi_jiffies;
 512                 }
 513         } else {
 514                 __cpufreq_driver_target(dbs_info->cur_policy,
 515                                         dbs_info->freq_lo,
 516                                         CPUFREQ_RELATION_H);
 517         }
 518         queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
 519         unlock_policy_rwsem_write(cpu);
 520 }
 521
 522 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
 523 {
 524         /* We want all CPUs to do sampling nearly on same jiffy */
 525         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 526         delay -= jiffies % delay;
 527
 528         dbs_info->enable = 1;
 529         ondemand_powersave_bias_init();
 530         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 531         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
 532         queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work,
 533                               delay);
 534 }
 535
 536 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
 537 {
 538         dbs_info->enable = 0;
 539         cancel_delayed_work(&dbs_info->work);
 540 }
 541
 542 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 543                                    unsigned int event)
 544 {
 545         unsigned int cpu = policy->cpu;
 546         struct cpu_dbs_info_s *this_dbs_info;
 547         unsigned int j;
 548         int rc;
 549
 550         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
 551
 552         switch (event) {
 553         case CPUFREQ_GOV_START:
 554                 if ((!cpu_online(cpu)) || (!policy->cur))
 555                         return -EINVAL;
 556
 557                 if (this_dbs_info->enable) /* Already enabled */
 558                         break;
 559
 560                 mutex_lock(&dbs_mutex);
 561                 dbs_enable++;
 562
 563                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
 564                 if (rc) {
 565                         dbs_enable--;
 566                         mutex_unlock(&dbs_mutex);
 567                         return rc;
 568                 }
 569
 570                 for_each_cpu(j, policy->cpus) {
 571                         struct cpu_dbs_info_s *j_dbs_info;
 572                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
 573                         j_dbs_info->cur_policy = policy;
 574
 575                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 576                                                 &j_dbs_info->prev_cpu_wall);
 577                         if (dbs_tuners_ins.ignore_nice) {
 578                                 j_dbs_info->prev_cpu_nice =
 579                                                 kstat_cpu(j).cpustat.nice;
 580                         }
 581                 }
 582                 this_dbs_info->cpu = cpu;
 583                 /*
 584                  * Start the timerschedule work, when this governor
 585                  * is used for first time
 586                  */
 587                 if (dbs_enable == 1) {
 588                         unsigned int latency;
 589                         /* policy latency is in nS. Convert it to uS first */
 590                         latency = policy->cpuinfo.transition_latency / 1000;
 591                         if (latency == 0)
 592                                 latency = 1;
 593
 594                         def_sampling_rate = latency *
 595                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
 596
 597                         if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
 598                                 def_sampling_rate = MIN_STAT_SAMPLING_RATE;
 599
 600                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
 601                 }
 602                 dbs_timer_init(this_dbs_info);
 603
 604                 mutex_unlock(&dbs_mutex);
 605                 break;
 606
 607         case CPUFREQ_GOV_STOP:
 608                 mutex_lock(&dbs_mutex);
 609                 dbs_timer_exit(this_dbs_info);
 610                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
 611                 dbs_enable--;
 612                 mutex_unlock(&dbs_mutex);
 613
 614                 break;
 615
 616         case CPUFREQ_GOV_LIMITS:
 617                 mutex_lock(&dbs_mutex);
 618                 if (policy->max < this_dbs_info->cur_policy->cur)
 619                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 620                                                 policy->max,
 621                                                 CPUFREQ_RELATION_H);
 622                 else if (policy->min > this_dbs_info->cur_policy->cur)
 623                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 624                                                 policy->min,
 625                                                 CPUFREQ_RELATION_L);
 626                 mutex_unlock(&dbs_mutex);
 627                 break;
 628         }
 629         return 0;
 630 }
 631
 632 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 633 static
 634 #endif
 635 struct cpufreq_governor cpufreq_gov_ondemand = {
 636         .name                   = "ondemand",
 637         .governor               = cpufreq_governor_dbs,
 638         .max_transition_latency = TRANSITION_LATENCY_LIMIT,
 639         .owner                  = THIS_MODULE,
 640 };
 641
 642 static int __init cpufreq_gov_dbs_init(void)
 643 {
 644         int err;
 645         cputime64_t wall;
 646         u64 idle_time;
 647         int cpu = get_cpu();
 648
 649         idle_time = get_cpu_idle_time_us(cpu, &wall);
 650         put_cpu();
 651         if (idle_time != -1ULL) {
 652                 /* Idle micro accounting is supported. Use finer thresholds */
 653                 dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
 654                 dbs_tuners_ins.down_differential =
 655                                         MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
 656         }
 657
 658         kondemand_wq = create_workqueue("kondemand");
 659         if (!kondemand_wq) {
 660                 printk(KERN_ERR "Creation of kondemand failed\n");
 661                 return -EFAULT;
 662         }
 663         err = cpufreq_register_governor(&cpufreq_gov_ondemand);
 664         if (err)
 665                 destroy_workqueue(kondemand_wq);
 666
 667         return err;
 668 }
 669
 670 static void __exit cpufreq_gov_dbs_exit(void)
 671 {
 672         cpufreq_unregister_governor(&cpufreq_gov_ondemand);
 673         destroy_workqueue(kondemand_wq);
 674 }
 675
 676
 677 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 678 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 679 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 680                    "Low Latency Frequency Transition capable processors");
 681 MODULE_LICENSE("GPL");
 682
 683 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 684 fs_initcall(cpufreq_gov_dbs_init);
 685 #else
 686 module_init(cpufreq_gov_dbs_init);
 687 #endif
 688 module_exit(cpufreq_gov_dbs_exit);