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 #include <linux/sched.h>
  25
  26 /*
  27  * dbs is used in this file as a shortform for demandbased switching
  28  * It helps to keep variable names smaller, simpler
  29  */
  30
  31 #define DEF_FREQUENCY_DOWN_DIFFERENTIAL         (10)
  32 #define DEF_FREQUENCY_UP_THRESHOLD              (80)
  33 #define MICRO_FREQUENCY_DOWN_DIFFERENTIAL       (3)
  34 #define MICRO_FREQUENCY_UP_THRESHOLD            (95)
  35 #define MICRO_FREQUENCY_MIN_SAMPLE_RATE         (10000)
  36 #define MIN_FREQUENCY_UP_THRESHOLD              (11)
  37 #define MAX_FREQUENCY_UP_THRESHOLD              (100)
  38
  39 /*
  40  * The polling frequency of this governor depends on the capability of
  41  * the processor. Default polling frequency is 1000 times the transition
  42  * latency of the processor. The governor will work on any processor with
  43  * transition latency <= 10mS, using appropriate sampling
  44  * rate.
  45  * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
  46  * this governor will not work.
  47  * All times here are in uS.
  48  */
  49 #define MIN_SAMPLING_RATE_RATIO                 (2)
  50
  51 static unsigned int min_sampling_rate;
  52
  53 #define LATENCY_MULTIPLIER                      (1000)
  54 #define MIN_LATENCY_MULTIPLIER                  (100)
  55 #define TRANSITION_LATENCY_LIMIT                (10 * 1000 * 1000)
  56
  57 static void do_dbs_timer(struct work_struct *work);
  58
  59 /* Sampling types */
  60 enum {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE};
  61
  62 struct cpu_dbs_info_s {
  63         cputime64_t prev_cpu_idle;
  64         cputime64_t prev_cpu_wall;
  65         cputime64_t prev_cpu_nice;
  66         struct cpufreq_policy *cur_policy;
  67         struct delayed_work work;
  68         struct cpufreq_frequency_table *freq_table;
  69         unsigned int freq_lo;
  70         unsigned int freq_lo_jiffies;
  71         unsigned int freq_hi_jiffies;
  72         int cpu;
  73         unsigned int sample_type:1;
  74         /*
  75          * percpu mutex that serializes governor limit change with
  76          * do_dbs_timer invocation. We do not want do_dbs_timer to run
  77          * when user is changing the governor or limits.
  78          */
  79         struct mutex timer_mutex;
  80 };
  81 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
  82
  83 static unsigned int dbs_enable; /* number of CPUs using this policy */
  84
  85 /*
  86  * dbs_mutex protects data in dbs_tuners_ins from concurrent changes on
  87  * different CPUs. It protects dbs_enable in governor start/stop.
  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_cpu(int cpu)
 194 {
 195         struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, cpu);
 196         dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
 197         dbs_info->freq_lo = 0;
 198 }
 199
 200 static void ondemand_powersave_bias_init(void)
 201 {
 202         int i;
 203         for_each_online_cpu(i) {
 204                 ondemand_powersave_bias_init_cpu(i);
 205         }
 206 }
 207
 208 /************************** sysfs interface ************************/
 209 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
 210 {
 211         printk_once(KERN_INFO "CPUFREQ: ondemand sampling_rate_max "
 212                "sysfs file is deprecated - used by: %s\n", current->comm);
 213         return sprintf(buf, "%u\n", -1U);
 214 }
 215
 216 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
 217 {
 218         return sprintf(buf, "%u\n", min_sampling_rate);
 219 }
 220
 221 #define define_one_ro(_name)            \
 222 static struct freq_attr _name =         \
 223 __ATTR(_name, 0444, show_##_name, NULL)
 224
 225 define_one_ro(sampling_rate_max);
 226 define_one_ro(sampling_rate_min);
 227
 228 /* cpufreq_ondemand Governor Tunables */
 229 #define show_one(file_name, object)                                     \
 230 static ssize_t show_##file_name                                         \
 231 (struct cpufreq_policy *unused, char *buf)                              \
 232 {                                                                       \
 233         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
 234 }
 235 show_one(sampling_rate, sampling_rate);
 236 show_one(up_threshold, up_threshold);
 237 show_one(ignore_nice_load, ignore_nice);
 238 show_one(powersave_bias, powersave_bias);
 239
 240 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
 241                 const char *buf, size_t count)
 242 {
 243         unsigned int input;
 244         int ret;
 245         ret = sscanf(buf, "%u", &input);
 246         if (ret != 1)
 247                 return -EINVAL;
 248
 249         mutex_lock(&dbs_mutex);
 250         dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
 251         mutex_unlock(&dbs_mutex);
 252
 253         return count;
 254 }
 255
 256 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
 257                 const char *buf, size_t count)
 258 {
 259         unsigned int input;
 260         int ret;
 261         ret = sscanf(buf, "%u", &input);
 262
 263         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 264                         input < MIN_FREQUENCY_UP_THRESHOLD) {
 265                 return -EINVAL;
 266         }
 267
 268         mutex_lock(&dbs_mutex);
 269         dbs_tuners_ins.up_threshold = input;
 270         mutex_unlock(&dbs_mutex);
 271
 272         return count;
 273 }
 274
 275 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
 276                 const char *buf, size_t count)
 277 {
 278         unsigned int input;
 279         int ret;
 280
 281         unsigned int j;
 282
 283         ret = sscanf(buf, "%u", &input);
 284         if (ret != 1)
 285                 return -EINVAL;
 286
 287         if (input > 1)
 288                 input = 1;
 289
 290         mutex_lock(&dbs_mutex);
 291         if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
 292                 mutex_unlock(&dbs_mutex);
 293                 return count;
 294         }
 295         dbs_tuners_ins.ignore_nice = input;
 296
 297         /* we need to re-evaluate prev_cpu_idle */
 298         for_each_online_cpu(j) {
 299                 struct cpu_dbs_info_s *dbs_info;
 300                 dbs_info = &per_cpu(cpu_dbs_info, j);
 301                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 302                                                 &dbs_info->prev_cpu_wall);
 303                 if (dbs_tuners_ins.ignore_nice)
 304                         dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 305
 306         }
 307         mutex_unlock(&dbs_mutex);
 308
 309         return count;
 310 }
 311
 312 static ssize_t store_powersave_bias(struct cpufreq_policy *unused,
 313                 const char *buf, size_t count)
 314 {
 315         unsigned int input;
 316         int ret;
 317         ret = sscanf(buf, "%u", &input);
 318
 319         if (ret != 1)
 320                 return -EINVAL;
 321
 322         if (input > 1000)
 323                 input = 1000;
 324
 325         mutex_lock(&dbs_mutex);
 326         dbs_tuners_ins.powersave_bias = input;
 327         ondemand_powersave_bias_init();
 328         mutex_unlock(&dbs_mutex);
 329
 330         return count;
 331 }
 332
 333 #define define_one_rw(_name) \
 334 static struct freq_attr _name = \
 335 __ATTR(_name, 0644, show_##_name, store_##_name)
 336
 337 define_one_rw(sampling_rate);
 338 define_one_rw(up_threshold);
 339 define_one_rw(ignore_nice_load);
 340 define_one_rw(powersave_bias);
 341
 342 static struct attribute *dbs_attributes[] = {
 343         &sampling_rate_max.attr,
 344         &sampling_rate_min.attr,
 345         &sampling_rate.attr,
 346         &up_threshold.attr,
 347         &ignore_nice_load.attr,
 348         &powersave_bias.attr,
 349         NULL
 350 };
 351
 352 static struct attribute_group dbs_attr_group = {
 353         .attrs = dbs_attributes,
 354         .name = "ondemand",
 355 };
 356
 357 /************************** sysfs end ************************/
 358
 359 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
 360 {
 361         unsigned int max_load_freq;
 362
 363         struct cpufreq_policy *policy;
 364         unsigned int j;
 365
 366         this_dbs_info->freq_lo = 0;
 367         policy = this_dbs_info->cur_policy;
 368
 369         /*
 370          * Every sampling_rate, we check, if current idle time is less
 371          * than 20% (default), then we try to increase frequency
 372          * Every sampling_rate, we look for a the lowest
 373          * frequency which can sustain the load while keeping idle time over
 374          * 30%. If such a frequency exist, we try to decrease to this frequency.
 375          *
 376          * Any frequency increase takes it to the maximum frequency.
 377          * Frequency reduction happens at minimum steps of
 378          * 5% (default) of current frequency
 379          */
 380
 381         /* Get Absolute Load - in terms of freq */
 382         max_load_freq = 0;
 383
 384         for_each_cpu(j, policy->cpus) {
 385                 struct cpu_dbs_info_s *j_dbs_info;
 386                 cputime64_t cur_wall_time, cur_idle_time;
 387                 unsigned int idle_time, wall_time;
 388                 unsigned int load, load_freq;
 389                 int freq_avg;
 390
 391                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
 392
 393                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
 394
 395                 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
 396                                 j_dbs_info->prev_cpu_wall);
 397                 j_dbs_info->prev_cpu_wall = cur_wall_time;
 398
 399                 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
 400                                 j_dbs_info->prev_cpu_idle);
 401                 j_dbs_info->prev_cpu_idle = cur_idle_time;
 402
 403                 if (dbs_tuners_ins.ignore_nice) {
 404                         cputime64_t cur_nice;
 405                         unsigned long cur_nice_jiffies;
 406
 407                         cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
 408                                          j_dbs_info->prev_cpu_nice);
 409                         /*
 410                          * Assumption: nice time between sampling periods will
 411                          * be less than 2^32 jiffies for 32 bit sys
 412                          */
 413                         cur_nice_jiffies = (unsigned long)
 414                                         cputime64_to_jiffies64(cur_nice);
 415
 416                         j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 417                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
 418                 }
 419
 420                 if (unlikely(!wall_time || wall_time < idle_time))
 421                         continue;
 422
 423                 load = 100 * (wall_time - idle_time) / wall_time;
 424
 425                 freq_avg = __cpufreq_driver_getavg(policy, j);
 426                 if (freq_avg <= 0)
 427                         freq_avg = policy->cur;
 428
 429                 load_freq = load * freq_avg;
 430                 if (load_freq > max_load_freq)
 431                         max_load_freq = load_freq;
 432         }
 433
 434         /* Check for frequency increase */
 435         if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
 436                 /* if we are already at full speed then break out early */
 437                 if (!dbs_tuners_ins.powersave_bias) {
 438                         if (policy->cur == policy->max)
 439                                 return;
 440
 441                         __cpufreq_driver_target(policy, policy->max,
 442                                 CPUFREQ_RELATION_H);
 443                 } else {
 444                         int freq = powersave_bias_target(policy, policy->max,
 445                                         CPUFREQ_RELATION_H);
 446                         __cpufreq_driver_target(policy, freq,
 447                                 CPUFREQ_RELATION_L);
 448                 }
 449                 return;
 450         }
 451
 452         /* Check for frequency decrease */
 453         /* if we cannot reduce the frequency anymore, break out early */
 454         if (policy->cur == policy->min)
 455                 return;
 456
 457         /*
 458          * The optimal frequency is the frequency that is the lowest that
 459          * can support the current CPU usage without triggering the up
 460          * policy. To be safe, we focus 10 points under the threshold.
 461          */
 462         if (max_load_freq <
 463             (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
 464              policy->cur) {
 465                 unsigned int freq_next;
 466                 freq_next = max_load_freq /
 467                                 (dbs_tuners_ins.up_threshold -
 468                                  dbs_tuners_ins.down_differential);
 469
 470                 if (!dbs_tuners_ins.powersave_bias) {
 471                         __cpufreq_driver_target(policy, freq_next,
 472                                         CPUFREQ_RELATION_L);
 473                 } else {
 474                         int freq = powersave_bias_target(policy, freq_next,
 475                                         CPUFREQ_RELATION_L);
 476                         __cpufreq_driver_target(policy, freq,
 477                                 CPUFREQ_RELATION_L);
 478                 }
 479         }
 480 }
 481
 482 static void do_dbs_timer(struct work_struct *work)
 483 {
 484         struct cpu_dbs_info_s *dbs_info =
 485                 container_of(work, struct cpu_dbs_info_s, work.work);
 486         unsigned int cpu = dbs_info->cpu;
 487         int sample_type = dbs_info->sample_type;
 488
 489         /* We want all CPUs to do sampling nearly on same jiffy */
 490         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 491
 492         delay -= jiffies % delay;
 493         mutex_lock(&dbs_info->timer_mutex);
 494
 495         /* Common NORMAL_SAMPLE setup */
 496         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 497         if (!dbs_tuners_ins.powersave_bias ||
 498             sample_type == DBS_NORMAL_SAMPLE) {
 499                 dbs_check_cpu(dbs_info);
 500                 if (dbs_info->freq_lo) {
 501                         /* Setup timer for SUB_SAMPLE */
 502                         dbs_info->sample_type = DBS_SUB_SAMPLE;
 503                         delay = dbs_info->freq_hi_jiffies;
 504                 }
 505         } else {
 506                 __cpufreq_driver_target(dbs_info->cur_policy,
 507                         dbs_info->freq_lo, CPUFREQ_RELATION_H);
 508         }
 509         queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
 510         mutex_unlock(&dbs_info->timer_mutex);
 511 }
 512
 513 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
 514 {
 515         /* We want all CPUs to do sampling nearly on same jiffy */
 516         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 517         delay -= jiffies % delay;
 518
 519         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 520         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
 521         queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work,
 522                 delay);
 523 }
 524
 525 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
 526 {
 527         cancel_delayed_work_sync(&dbs_info->work);
 528 }
 529
 530 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 531                                    unsigned int event)
 532 {
 533         unsigned int cpu = policy->cpu;
 534         struct cpu_dbs_info_s *this_dbs_info;
 535         unsigned int j;
 536         int rc;
 537
 538         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
 539
 540         switch (event) {
 541         case CPUFREQ_GOV_START:
 542                 if ((!cpu_online(cpu)) || (!policy->cur))
 543                         return -EINVAL;
 544
 545                 mutex_lock(&dbs_mutex);
 546
 547                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
 548                 if (rc) {
 549                         mutex_unlock(&dbs_mutex);
 550                         return rc;
 551                 }
 552
 553                 dbs_enable++;
 554                 for_each_cpu(j, policy->cpus) {
 555                         struct cpu_dbs_info_s *j_dbs_info;
 556                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
 557                         j_dbs_info->cur_policy = policy;
 558
 559                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 560                                                 &j_dbs_info->prev_cpu_wall);
 561                         if (dbs_tuners_ins.ignore_nice) {
 562                                 j_dbs_info->prev_cpu_nice =
 563                                                 kstat_cpu(j).cpustat.nice;
 564                         }
 565                 }
 566                 this_dbs_info->cpu = cpu;
 567                 ondemand_powersave_bias_init_cpu(cpu);
 568                 mutex_init(&this_dbs_info->timer_mutex);
 569                 /*
 570                  * Start the timerschedule work, when this governor
 571                  * is used for first time
 572                  */
 573                 if (dbs_enable == 1) {
 574                         unsigned int latency;
 575                         /* policy latency is in nS. Convert it to uS first */
 576                         latency = policy->cpuinfo.transition_latency / 1000;
 577                         if (latency == 0)
 578                                 latency = 1;
 579                         /* Bring kernel and HW constraints together */
 580                         min_sampling_rate = max(min_sampling_rate,
 581                                         MIN_LATENCY_MULTIPLIER * latency);
 582                         dbs_tuners_ins.sampling_rate =
 583                                 max(min_sampling_rate,
 584                                     latency * LATENCY_MULTIPLIER);
 585                 }
 586                 mutex_unlock(&dbs_mutex);
 587
 588                 dbs_timer_init(this_dbs_info);
 589                 break;
 590
 591         case CPUFREQ_GOV_STOP:
 592                 dbs_timer_exit(this_dbs_info);
 593
 594                 mutex_lock(&dbs_mutex);
 595                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
 596                 mutex_destroy(&this_dbs_info->timer_mutex);
 597                 dbs_enable--;
 598                 mutex_unlock(&dbs_mutex);
 599
 600                 break;
 601
 602         case CPUFREQ_GOV_LIMITS:
 603                 mutex_lock(&this_dbs_info->timer_mutex);
 604                 if (policy->max < this_dbs_info->cur_policy->cur)
 605                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 606                                 policy->max, CPUFREQ_RELATION_H);
 607                 else if (policy->min > this_dbs_info->cur_policy->cur)
 608                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 609                                 policy->min, CPUFREQ_RELATION_L);
 610                 mutex_unlock(&this_dbs_info->timer_mutex);
 611                 break;
 612         }
 613         return 0;
 614 }
 615
 616 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 617 static
 618 #endif
 619 struct cpufreq_governor cpufreq_gov_ondemand = {
 620         .name                   = "ondemand",
 621         .governor               = cpufreq_governor_dbs,
 622         .max_transition_latency = TRANSITION_LATENCY_LIMIT,
 623         .owner                  = THIS_MODULE,
 624 };
 625
 626 static int __init cpufreq_gov_dbs_init(void)
 627 {
 628         int err;
 629         cputime64_t wall;
 630         u64 idle_time;
 631         int cpu = get_cpu();
 632
 633         idle_time = get_cpu_idle_time_us(cpu, &wall);
 634         put_cpu();
 635         if (idle_time != -1ULL) {
 636                 /* Idle micro accounting is supported. Use finer thresholds */
 637                 dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
 638                 dbs_tuners_ins.down_differential =
 639                                         MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
 640                 /*
 641                  * In no_hz/micro accounting case we set the minimum frequency
 642                  * not depending on HZ, but fixed (very low). The deferred
 643                  * timer might skip some samples if idle/sleeping as needed.
 644                 */
 645                 min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
 646         } else {
 647                 /* For correct statistics, we need 10 ticks for each measure */
 648                 min_sampling_rate =
 649                         MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
 650         }
 651
 652         kondemand_wq = create_workqueue("kondemand");
 653         if (!kondemand_wq) {
 654                 printk(KERN_ERR "Creation of kondemand failed\n");
 655                 return -EFAULT;
 656         }
 657         err = cpufreq_register_governor(&cpufreq_gov_ondemand);
 658         if (err)
 659                 destroy_workqueue(kondemand_wq);
 660
 661         return err;
 662 }
 663
 664 static void __exit cpufreq_gov_dbs_exit(void)
 665 {
 666         cpufreq_unregister_governor(&cpufreq_gov_ondemand);
 667         destroy_workqueue(kondemand_wq);
 668 }
 669
 670
 671 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 672 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 673 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 674         "Low Latency Frequency Transition capable processors");
 675 MODULE_LICENSE("GPL");
 676
 677 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 678 fs_initcall(cpufreq_gov_dbs_init);
 679 #else
 680 module_init(cpufreq_gov_dbs_init);
 681 #endif
 682 module_exit(cpufreq_gov_dbs_exit);