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, od_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(od_cpu_dbs_info,
 153                                                    policy->cpu);
 154
 155         if (!dbs_info->freq_table) {
 156                 dbs_info->freq_lo = 0;
 157                 dbs_info->freq_lo_jiffies = 0;
 158                 return freq_next;
 159         }
 160
 161         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
 162                         relation, &index);
 163         freq_req = dbs_info->freq_table[index].frequency;
 164         freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
 165         freq_avg = freq_req - freq_reduc;
 166
 167         /* Find freq bounds for freq_avg in freq_table */
 168         index = 0;
 169         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 170                         CPUFREQ_RELATION_H, &index);
 171         freq_lo = dbs_info->freq_table[index].frequency;
 172         index = 0;
 173         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 174                         CPUFREQ_RELATION_L, &index);
 175         freq_hi = dbs_info->freq_table[index].frequency;
 176
 177         /* Find out how long we have to be in hi and lo freqs */
 178         if (freq_hi == freq_lo) {
 179                 dbs_info->freq_lo = 0;
 180                 dbs_info->freq_lo_jiffies = 0;
 181                 return freq_lo;
 182         }
 183         jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 184         jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
 185         jiffies_hi += ((freq_hi - freq_lo) / 2);
 186         jiffies_hi /= (freq_hi - freq_lo);
 187         jiffies_lo = jiffies_total - jiffies_hi;
 188         dbs_info->freq_lo = freq_lo;
 189         dbs_info->freq_lo_jiffies = jiffies_lo;
 190         dbs_info->freq_hi_jiffies = jiffies_hi;
 191         return freq_hi;
 192 }
 193
 194 static void ondemand_powersave_bias_init_cpu(int cpu)
 195 {
 196         struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 197         dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
 198         dbs_info->freq_lo = 0;
 199 }
 200
 201 static void ondemand_powersave_bias_init(void)
 202 {
 203         int i;
 204         for_each_online_cpu(i) {
 205                 ondemand_powersave_bias_init_cpu(i);
 206         }
 207 }
 208
 209 /************************** sysfs interface ************************/
 210 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
 211 {
 212         printk_once(KERN_INFO "CPUFREQ: ondemand sampling_rate_max "
 213                "sysfs file is deprecated - used by: %s\n", current->comm);
 214         return sprintf(buf, "%u\n", -1U);
 215 }
 216
 217 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
 218 {
 219         return sprintf(buf, "%u\n", min_sampling_rate);
 220 }
 221
 222 #define define_one_ro(_name)            \
 223 static struct freq_attr _name =         \
 224 __ATTR(_name, 0444, show_##_name, NULL)
 225
 226 define_one_ro(sampling_rate_max);
 227 define_one_ro(sampling_rate_min);
 228
 229 /* cpufreq_ondemand Governor Tunables */
 230 #define show_one(file_name, object)                                     \
 231 static ssize_t show_##file_name                                         \
 232 (struct cpufreq_policy *unused, char *buf)                              \
 233 {                                                                       \
 234         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
 235 }
 236 show_one(sampling_rate, sampling_rate);
 237 show_one(up_threshold, up_threshold);
 238 show_one(ignore_nice_load, ignore_nice);
 239 show_one(powersave_bias, powersave_bias);
 240
 241 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
 242                 const char *buf, size_t count)
 243 {
 244         unsigned int input;
 245         int ret;
 246         ret = sscanf(buf, "%u", &input);
 247         if (ret != 1)
 248                 return -EINVAL;
 249
 250         mutex_lock(&dbs_mutex);
 251         dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
 252         mutex_unlock(&dbs_mutex);
 253
 254         return count;
 255 }
 256
 257 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
 258                 const char *buf, size_t count)
 259 {
 260         unsigned int input;
 261         int ret;
 262         ret = sscanf(buf, "%u", &input);
 263
 264         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 265                         input < MIN_FREQUENCY_UP_THRESHOLD) {
 266                 return -EINVAL;
 267         }
 268
 269         mutex_lock(&dbs_mutex);
 270         dbs_tuners_ins.up_threshold = input;
 271         mutex_unlock(&dbs_mutex);
 272
 273         return count;
 274 }
 275
 276 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
 277                 const char *buf, size_t count)
 278 {
 279         unsigned int input;
 280         int ret;
 281
 282         unsigned int j;
 283
 284         ret = sscanf(buf, "%u", &input);
 285         if (ret != 1)
 286                 return -EINVAL;
 287
 288         if (input > 1)
 289                 input = 1;
 290
 291         mutex_lock(&dbs_mutex);
 292         if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
 293                 mutex_unlock(&dbs_mutex);
 294                 return count;
 295         }
 296         dbs_tuners_ins.ignore_nice = input;
 297
 298         /* we need to re-evaluate prev_cpu_idle */
 299         for_each_online_cpu(j) {
 300                 struct cpu_dbs_info_s *dbs_info;
 301                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
 302                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 303                                                 &dbs_info->prev_cpu_wall);
 304                 if (dbs_tuners_ins.ignore_nice)
 305                         dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 306
 307         }
 308         mutex_unlock(&dbs_mutex);
 309
 310         return count;
 311 }
 312
 313 static ssize_t store_powersave_bias(struct cpufreq_policy *unused,
 314                 const char *buf, size_t count)
 315 {
 316         unsigned int input;
 317         int ret;
 318         ret = sscanf(buf, "%u", &input);
 319
 320         if (ret != 1)
 321                 return -EINVAL;
 322
 323         if (input > 1000)
 324                 input = 1000;
 325
 326         mutex_lock(&dbs_mutex);
 327         dbs_tuners_ins.powersave_bias = input;
 328         ondemand_powersave_bias_init();
 329         mutex_unlock(&dbs_mutex);
 330
 331         return count;
 332 }
 333
 334 #define define_one_rw(_name) \
 335 static struct freq_attr _name = \
 336 __ATTR(_name, 0644, show_##_name, store_##_name)
 337
 338 define_one_rw(sampling_rate);
 339 define_one_rw(up_threshold);
 340 define_one_rw(ignore_nice_load);
 341 define_one_rw(powersave_bias);
 342
 343 static struct attribute *dbs_attributes[] = {
 344         &sampling_rate_max.attr,
 345         &sampling_rate_min.attr,
 346         &sampling_rate.attr,
 347         &up_threshold.attr,
 348         &ignore_nice_load.attr,
 349         &powersave_bias.attr,
 350         NULL
 351 };
 352
 353 static struct attribute_group dbs_attr_group = {
 354         .attrs = dbs_attributes,
 355         .name = "ondemand",
 356 };
 357
 358 /************************** sysfs end ************************/
 359
 360 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
 361 {
 362         unsigned int max_load_freq;
 363
 364         struct cpufreq_policy *policy;
 365         unsigned int j;
 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(od_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         mutex_lock(&dbs_info->timer_mutex);
 495
 496         /* Common NORMAL_SAMPLE setup */
 497         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 498         if (!dbs_tuners_ins.powersave_bias ||
 499             sample_type == DBS_NORMAL_SAMPLE) {
 500                 dbs_check_cpu(dbs_info);
 501                 if (dbs_info->freq_lo) {
 502                         /* Setup timer for SUB_SAMPLE */
 503                         dbs_info->sample_type = DBS_SUB_SAMPLE;
 504                         delay = dbs_info->freq_hi_jiffies;
 505                 }
 506         } else {
 507                 __cpufreq_driver_target(dbs_info->cur_policy,
 508                         dbs_info->freq_lo, CPUFREQ_RELATION_H);
 509         }
 510         queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
 511         mutex_unlock(&dbs_info->timer_mutex);
 512 }
 513
 514 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
 515 {
 516         /* We want all CPUs to do sampling nearly on same jiffy */
 517         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 518         delay -= jiffies % delay;
 519
 520         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 521         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
 522         queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work,
 523                 delay);
 524 }
 525
 526 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
 527 {
 528         cancel_delayed_work_sync(&dbs_info->work);
 529 }
 530
 531 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 532                                    unsigned int event)
 533 {
 534         unsigned int cpu = policy->cpu;
 535         struct cpu_dbs_info_s *this_dbs_info;
 536         unsigned int j;
 537         int rc;
 538
 539         this_dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 540
 541         switch (event) {
 542         case CPUFREQ_GOV_START:
 543                 if ((!cpu_online(cpu)) || (!policy->cur))
 544                         return -EINVAL;
 545
 546                 mutex_lock(&dbs_mutex);
 547
 548                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
 549                 if (rc) {
 550                         mutex_unlock(&dbs_mutex);
 551                         return rc;
 552                 }
 553
 554                 dbs_enable++;
 555                 for_each_cpu(j, policy->cpus) {
 556                         struct cpu_dbs_info_s *j_dbs_info;
 557                         j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
 558                         j_dbs_info->cur_policy = policy;
 559
 560                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 561                                                 &j_dbs_info->prev_cpu_wall);
 562                         if (dbs_tuners_ins.ignore_nice) {
 563                                 j_dbs_info->prev_cpu_nice =
 564                                                 kstat_cpu(j).cpustat.nice;
 565                         }
 566                 }
 567                 this_dbs_info->cpu = cpu;
 568                 ondemand_powersave_bias_init_cpu(cpu);
 569                 mutex_init(&this_dbs_info->timer_mutex);
 570                 /*
 571                  * Start the timerschedule work, when this governor
 572                  * is used for first time
 573                  */
 574                 if (dbs_enable == 1) {
 575                         unsigned int latency;
 576                         /* policy latency is in nS. Convert it to uS first */
 577                         latency = policy->cpuinfo.transition_latency / 1000;
 578                         if (latency == 0)
 579                                 latency = 1;
 580                         /* Bring kernel and HW constraints together */
 581                         min_sampling_rate = max(min_sampling_rate,
 582                                         MIN_LATENCY_MULTIPLIER * latency);
 583                         dbs_tuners_ins.sampling_rate =
 584                                 max(min_sampling_rate,
 585                                     latency * LATENCY_MULTIPLIER);
 586                 }
 587                 mutex_unlock(&dbs_mutex);
 588
 589                 dbs_timer_init(this_dbs_info);
 590                 break;
 591
 592         case CPUFREQ_GOV_STOP:
 593                 dbs_timer_exit(this_dbs_info);
 594
 595                 mutex_lock(&dbs_mutex);
 596                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
 597                 mutex_destroy(&this_dbs_info->timer_mutex);
 598                 dbs_enable--;
 599                 mutex_unlock(&dbs_mutex);
 600
 601                 break;
 602
 603         case CPUFREQ_GOV_LIMITS:
 604                 mutex_lock(&this_dbs_info->timer_mutex);
 605                 if (policy->max < this_dbs_info->cur_policy->cur)
 606                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 607                                 policy->max, CPUFREQ_RELATION_H);
 608                 else if (policy->min > this_dbs_info->cur_policy->cur)
 609                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 610                                 policy->min, CPUFREQ_RELATION_L);
 611                 mutex_unlock(&this_dbs_info->timer_mutex);
 612                 break;
 613         }
 614         return 0;
 615 }
 616
 617 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 618 static
 619 #endif
 620 struct cpufreq_governor cpufreq_gov_ondemand = {
 621         .name                   = "ondemand",
 622         .governor               = cpufreq_governor_dbs,
 623         .max_transition_latency = TRANSITION_LATENCY_LIMIT,
 624         .owner                  = THIS_MODULE,
 625 };
 626
 627 static int __init cpufreq_gov_dbs_init(void)
 628 {
 629         int err;
 630         cputime64_t wall;
 631         u64 idle_time;
 632         int cpu = get_cpu();
 633
 634         idle_time = get_cpu_idle_time_us(cpu, &wall);
 635         put_cpu();
 636         if (idle_time != -1ULL) {
 637                 /* Idle micro accounting is supported. Use finer thresholds */
 638                 dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
 639                 dbs_tuners_ins.down_differential =
 640                                         MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
 641                 /*
 642                  * In no_hz/micro accounting case we set the minimum frequency
 643                  * not depending on HZ, but fixed (very low). The deferred
 644                  * timer might skip some samples if idle/sleeping as needed.
 645                 */
 646                 min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
 647         } else {
 648                 /* For correct statistics, we need 10 ticks for each measure */
 649                 min_sampling_rate =
 650                         MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
 651         }
 652
 653         kondemand_wq = create_workqueue("kondemand");
 654         if (!kondemand_wq) {
 655                 printk(KERN_ERR "Creation of kondemand failed\n");
 656                 return -EFAULT;
 657         }
 658         err = cpufreq_register_governor(&cpufreq_gov_ondemand);
 659         if (err)
 660                 destroy_workqueue(kondemand_wq);
 661
 662         return err;
 663 }
 664
 665 static void __exit cpufreq_gov_dbs_exit(void)
 666 {
 667         cpufreq_unregister_governor(&cpufreq_gov_ondemand);
 668         destroy_workqueue(kondemand_wq);
 669 }
 670
 671
 672 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 673 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 674 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 675         "Low Latency Frequency Transition capable processors");
 676 MODULE_LICENSE("GPL");
 677
 678 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 679 fs_initcall(cpufreq_gov_dbs_init);
 680 #else
 681 module_init(cpufreq_gov_dbs_init);
 682 #endif
 683 module_exit(cpufreq_gov_dbs_exit);