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 MIN_FREQUENCY_UP_THRESHOLD              (11)
  36 #define MAX_FREQUENCY_UP_THRESHOLD              (100)
  37
  38 /*
  39  * The polling frequency of this governor depends on the capability of
  40  * the processor. Default polling frequency is 1000 times the transition
  41  * latency of the processor. The governor will work on any processor with
  42  * transition latency <= 10mS, using appropriate sampling
  43  * rate.
  44  * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
  45  * this governor will not work.
  46  * All times here are in uS.
  47  */
  48 static unsigned int def_sampling_rate;
  49 #define MIN_SAMPLING_RATE_RATIO                 (2)
  50 /* for correct statistics, we need at least 10 ticks between each measure */
  51 #define MIN_STAT_SAMPLING_RATE                  \
  52                         (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
  53 #define MIN_SAMPLING_RATE                       \
  54                         (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
  55 /* Above MIN_SAMPLING_RATE will vanish with its sysfs file soon
  56  * Define the minimal settable sampling rate to the greater of:
  57  *   - "HW transition latency" * 100 (same as default sampling / 10)
  58  *   - MIN_STAT_SAMPLING_RATE
  59  * To avoid that userspace shoots itself.
  60 */
  61 static unsigned int minimum_sampling_rate(void)
  62 {
  63         return max(def_sampling_rate / 10, MIN_STAT_SAMPLING_RATE);
  64 }
  65
  66 /* This will also vanish soon with removing sampling_rate_max */
  67 #define MAX_SAMPLING_RATE                       (500 * def_sampling_rate)
  68 #define LATENCY_MULTIPLIER                      (1000)
  69 #define TRANSITION_LATENCY_LIMIT                (10 * 1000 * 1000)
  70
  71 static void do_dbs_timer(struct work_struct *work);
  72
  73 /* Sampling types */
  74 enum {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE};
  75
  76 struct cpu_dbs_info_s {
  77         cputime64_t prev_cpu_idle;
  78         cputime64_t prev_cpu_wall;
  79         cputime64_t prev_cpu_nice;
  80         struct cpufreq_policy *cur_policy;
  81         struct delayed_work work;
  82         struct cpufreq_frequency_table *freq_table;
  83         unsigned int freq_lo;
  84         unsigned int freq_lo_jiffies;
  85         unsigned int freq_hi_jiffies;
  86         int cpu;
  87         unsigned int enable:1,
  88                 sample_type:1;
  89 };
  90 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
  91
  92 static unsigned int dbs_enable; /* number of CPUs using this policy */
  93
  94 /*
  95  * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug
  96  * lock and dbs_mutex. cpu_hotplug lock should always be held before
  97  * dbs_mutex. If any function that can potentially take cpu_hotplug lock
  98  * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
  99  * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
 100  * is recursive for the same process. -Venki
 101  * DEADLOCK ALERT! (2) : do_dbs_timer() must not take the dbs_mutex, because it
 102  * would deadlock with cancel_delayed_work_sync(), which is needed for proper
 103  * raceless workqueue teardown.
 104  */
 105 static DEFINE_MUTEX(dbs_mutex);
 106
 107 static struct workqueue_struct  *kondemand_wq;
 108
 109 static struct dbs_tuners {
 110         unsigned int sampling_rate;
 111         unsigned int up_threshold;
 112         unsigned int down_differential;
 113         unsigned int ignore_nice;
 114         unsigned int powersave_bias;
 115 } dbs_tuners_ins = {
 116         .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
 117         .down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL,
 118         .ignore_nice = 0,
 119         .powersave_bias = 0,
 120 };
 121
 122 static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
 123                                                         cputime64_t *wall)
 124 {
 125         cputime64_t idle_time;
 126         cputime64_t cur_wall_time;
 127         cputime64_t busy_time;
 128
 129         cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
 130         busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
 131                         kstat_cpu(cpu).cpustat.system);
 132
 133         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
 134         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
 135         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
 136         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice);
 137
 138         idle_time = cputime64_sub(cur_wall_time, busy_time);
 139         if (wall)
 140                 *wall = cur_wall_time;
 141
 142         return idle_time;
 143 }
 144
 145 static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
 146 {
 147         u64 idle_time = get_cpu_idle_time_us(cpu, wall);
 148
 149         if (idle_time == -1ULL)
 150                 return get_cpu_idle_time_jiffy(cpu, wall);
 151
 152         return idle_time;
 153 }
 154
 155 /*
 156  * Find right freq to be set now with powersave_bias on.
 157  * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
 158  * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
 159  */
 160 static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
 161                                           unsigned int freq_next,
 162                                           unsigned int relation)
 163 {
 164         unsigned int freq_req, freq_reduc, freq_avg;
 165         unsigned int freq_hi, freq_lo;
 166         unsigned int index = 0;
 167         unsigned int jiffies_total, jiffies_hi, jiffies_lo;
 168         struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, policy->cpu);
 169
 170         if (!dbs_info->freq_table) {
 171                 dbs_info->freq_lo = 0;
 172                 dbs_info->freq_lo_jiffies = 0;
 173                 return freq_next;
 174         }
 175
 176         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
 177                         relation, &index);
 178         freq_req = dbs_info->freq_table[index].frequency;
 179         freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
 180         freq_avg = freq_req - freq_reduc;
 181
 182         /* Find freq bounds for freq_avg in freq_table */
 183         index = 0;
 184         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 185                         CPUFREQ_RELATION_H, &index);
 186         freq_lo = dbs_info->freq_table[index].frequency;
 187         index = 0;
 188         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 189                         CPUFREQ_RELATION_L, &index);
 190         freq_hi = dbs_info->freq_table[index].frequency;
 191
 192         /* Find out how long we have to be in hi and lo freqs */
 193         if (freq_hi == freq_lo) {
 194                 dbs_info->freq_lo = 0;
 195                 dbs_info->freq_lo_jiffies = 0;
 196                 return freq_lo;
 197         }
 198         jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 199         jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
 200         jiffies_hi += ((freq_hi - freq_lo) / 2);
 201         jiffies_hi /= (freq_hi - freq_lo);
 202         jiffies_lo = jiffies_total - jiffies_hi;
 203         dbs_info->freq_lo = freq_lo;
 204         dbs_info->freq_lo_jiffies = jiffies_lo;
 205         dbs_info->freq_hi_jiffies = jiffies_hi;
 206         return freq_hi;
 207 }
 208
 209 static void ondemand_powersave_bias_init(void)
 210 {
 211         int i;
 212         for_each_online_cpu(i) {
 213                 struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, i);
 214                 dbs_info->freq_table = cpufreq_frequency_get_table(i);
 215                 dbs_info->freq_lo = 0;
 216         }
 217 }
 218
 219 /************************** sysfs interface ************************/
 220 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
 221 {
 222         static int print_once;
 223
 224         if (!print_once) {
 225                 printk(KERN_INFO "CPUFREQ: ondemand sampling_rate_max "
 226                        "sysfs file is deprecated - used by: %s\n",
 227                        current->comm);
 228                 print_once = 1;
 229         }
 230         return sprintf(buf, "%u\n", MAX_SAMPLING_RATE);
 231 }
 232
 233 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
 234 {
 235         static int print_once;
 236
 237         if (!print_once) {
 238                 printk(KERN_INFO "CPUFREQ: ondemand sampling_rate_min "
 239                        "sysfs file is deprecated - used by: %s\n",
 240                        current->comm);
 241                 print_once = 1;
 242         }
 243         return sprintf(buf, "%u\n", MIN_SAMPLING_RATE);
 244 }
 245
 246 #define define_one_ro(_name)            \
 247 static struct freq_attr _name =         \
 248 __ATTR(_name, 0444, show_##_name, NULL)
 249
 250 define_one_ro(sampling_rate_max);
 251 define_one_ro(sampling_rate_min);
 252
 253 /* cpufreq_ondemand Governor Tunables */
 254 #define show_one(file_name, object)                                     \
 255 static ssize_t show_##file_name                                         \
 256 (struct cpufreq_policy *unused, char *buf)                              \
 257 {                                                                       \
 258         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
 259 }
 260 show_one(sampling_rate, sampling_rate);
 261 show_one(up_threshold, up_threshold);
 262 show_one(ignore_nice_load, ignore_nice);
 263 show_one(powersave_bias, powersave_bias);
 264
 265 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
 266                 const char *buf, size_t count)
 267 {
 268         unsigned int input;
 269         int ret;
 270         ret = sscanf(buf, "%u", &input);
 271
 272         mutex_lock(&dbs_mutex);
 273         if (ret != 1) {
 274                 mutex_unlock(&dbs_mutex);
 275                 return -EINVAL;
 276         }
 277         dbs_tuners_ins.sampling_rate = max(input, minimum_sampling_rate());
 278         mutex_unlock(&dbs_mutex);
 279
 280         return count;
 281 }
 282
 283 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
 284                 const char *buf, size_t count)
 285 {
 286         unsigned int input;
 287         int ret;
 288         ret = sscanf(buf, "%u", &input);
 289
 290         mutex_lock(&dbs_mutex);
 291         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 292                         input < MIN_FREQUENCY_UP_THRESHOLD) {
 293                 mutex_unlock(&dbs_mutex);
 294                 return -EINVAL;
 295         }
 296
 297         dbs_tuners_ins.up_threshold = input;
 298         mutex_unlock(&dbs_mutex);
 299
 300         return count;
 301 }
 302
 303 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
 304                 const char *buf, size_t count)
 305 {
 306         unsigned int input;
 307         int ret;
 308
 309         unsigned int j;
 310
 311         ret = sscanf(buf, "%u", &input);
 312         if (ret != 1)
 313                 return -EINVAL;
 314
 315         if (input > 1)
 316                 input = 1;
 317
 318         mutex_lock(&dbs_mutex);
 319         if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
 320                 mutex_unlock(&dbs_mutex);
 321                 return count;
 322         }
 323         dbs_tuners_ins.ignore_nice = input;
 324
 325         /* we need to re-evaluate prev_cpu_idle */
 326         for_each_online_cpu(j) {
 327                 struct cpu_dbs_info_s *dbs_info;
 328                 dbs_info = &per_cpu(cpu_dbs_info, j);
 329                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 330                                                 &dbs_info->prev_cpu_wall);
 331                 if (dbs_tuners_ins.ignore_nice)
 332                         dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 333
 334         }
 335         mutex_unlock(&dbs_mutex);
 336
 337         return count;
 338 }
 339
 340 static ssize_t store_powersave_bias(struct cpufreq_policy *unused,
 341                 const char *buf, size_t count)
 342 {
 343         unsigned int input;
 344         int ret;
 345         ret = sscanf(buf, "%u", &input);
 346
 347         if (ret != 1)
 348                 return -EINVAL;
 349
 350         if (input > 1000)
 351                 input = 1000;
 352
 353         mutex_lock(&dbs_mutex);
 354         dbs_tuners_ins.powersave_bias = input;
 355         ondemand_powersave_bias_init();
 356         mutex_unlock(&dbs_mutex);
 357
 358         return count;
 359 }
 360
 361 #define define_one_rw(_name) \
 362 static struct freq_attr _name = \
 363 __ATTR(_name, 0644, show_##_name, store_##_name)
 364
 365 define_one_rw(sampling_rate);
 366 define_one_rw(up_threshold);
 367 define_one_rw(ignore_nice_load);
 368 define_one_rw(powersave_bias);
 369
 370 static struct attribute *dbs_attributes[] = {
 371         &sampling_rate_max.attr,
 372         &sampling_rate_min.attr,
 373         &sampling_rate.attr,
 374         &up_threshold.attr,
 375         &ignore_nice_load.attr,
 376         &powersave_bias.attr,
 377         NULL
 378 };
 379
 380 static struct attribute_group dbs_attr_group = {
 381         .attrs = dbs_attributes,
 382         .name = "ondemand",
 383 };
 384
 385 /************************** sysfs end ************************/
 386
 387 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
 388 {
 389         unsigned int max_load_freq;
 390
 391         struct cpufreq_policy *policy;
 392         unsigned int j;
 393
 394         if (!this_dbs_info->enable)
 395                 return;
 396
 397         this_dbs_info->freq_lo = 0;
 398         policy = this_dbs_info->cur_policy;
 399
 400         /*
 401          * Every sampling_rate, we check, if current idle time is less
 402          * than 20% (default), then we try to increase frequency
 403          * Every sampling_rate, we look for a the lowest
 404          * frequency which can sustain the load while keeping idle time over
 405          * 30%. If such a frequency exist, we try to decrease to this frequency.
 406          *
 407          * Any frequency increase takes it to the maximum frequency.
 408          * Frequency reduction happens at minimum steps of
 409          * 5% (default) of current frequency
 410          */
 411
 412         /* Get Absolute Load - in terms of freq */
 413         max_load_freq = 0;
 414
 415         for_each_cpu(j, policy->cpus) {
 416                 struct cpu_dbs_info_s *j_dbs_info;
 417                 cputime64_t cur_wall_time, cur_idle_time;
 418                 unsigned int idle_time, wall_time;
 419                 unsigned int load, load_freq;
 420                 int freq_avg;
 421
 422                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
 423
 424                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
 425
 426                 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
 427                                 j_dbs_info->prev_cpu_wall);
 428                 j_dbs_info->prev_cpu_wall = cur_wall_time;
 429
 430                 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
 431                                 j_dbs_info->prev_cpu_idle);
 432                 j_dbs_info->prev_cpu_idle = cur_idle_time;
 433
 434                 if (dbs_tuners_ins.ignore_nice) {
 435                         cputime64_t cur_nice;
 436                         unsigned long cur_nice_jiffies;
 437
 438                         cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
 439                                          j_dbs_info->prev_cpu_nice);
 440                         /*
 441                          * Assumption: nice time between sampling periods will
 442                          * be less than 2^32 jiffies for 32 bit sys
 443                          */
 444                         cur_nice_jiffies = (unsigned long)
 445                                         cputime64_to_jiffies64(cur_nice);
 446
 447                         j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 448                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
 449                 }
 450
 451                 if (unlikely(!wall_time || wall_time < idle_time))
 452                         continue;
 453
 454                 load = 100 * (wall_time - idle_time) / wall_time;
 455
 456                 freq_avg = __cpufreq_driver_getavg(policy, j);
 457                 if (freq_avg <= 0)
 458                         freq_avg = policy->cur;
 459
 460                 load_freq = load * freq_avg;
 461                 if (load_freq > max_load_freq)
 462                         max_load_freq = load_freq;
 463         }
 464
 465         /* Check for frequency increase */
 466         if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
 467                 /* if we are already at full speed then break out early */
 468                 if (!dbs_tuners_ins.powersave_bias) {
 469                         if (policy->cur == policy->max)
 470                                 return;
 471
 472                         __cpufreq_driver_target(policy, policy->max,
 473                                 CPUFREQ_RELATION_H);
 474                 } else {
 475                         int freq = powersave_bias_target(policy, policy->max,
 476                                         CPUFREQ_RELATION_H);
 477                         __cpufreq_driver_target(policy, freq,
 478                                 CPUFREQ_RELATION_L);
 479                 }
 480                 return;
 481         }
 482
 483         /* Check for frequency decrease */
 484         /* if we cannot reduce the frequency anymore, break out early */
 485         if (policy->cur == policy->min)
 486                 return;
 487
 488         /*
 489          * The optimal frequency is the frequency that is the lowest that
 490          * can support the current CPU usage without triggering the up
 491          * policy. To be safe, we focus 10 points under the threshold.
 492          */
 493         if (max_load_freq <
 494             (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
 495              policy->cur) {
 496                 unsigned int freq_next;
 497                 freq_next = max_load_freq /
 498                                 (dbs_tuners_ins.up_threshold -
 499                                  dbs_tuners_ins.down_differential);
 500
 501                 if (!dbs_tuners_ins.powersave_bias) {
 502                         __cpufreq_driver_target(policy, freq_next,
 503                                         CPUFREQ_RELATION_L);
 504                 } else {
 505                         int freq = powersave_bias_target(policy, freq_next,
 506                                         CPUFREQ_RELATION_L);
 507                         __cpufreq_driver_target(policy, freq,
 508                                 CPUFREQ_RELATION_L);
 509                 }
 510         }
 511 }
 512
 513 static void do_dbs_timer(struct work_struct *work)
 514 {
 515         struct cpu_dbs_info_s *dbs_info =
 516                 container_of(work, struct cpu_dbs_info_s, work.work);
 517         unsigned int cpu = dbs_info->cpu;
 518         int sample_type = dbs_info->sample_type;
 519
 520         /* We want all CPUs to do sampling nearly on same jiffy */
 521         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 522
 523         delay -= jiffies % delay;
 524
 525         if (lock_policy_rwsem_write(cpu) < 0)
 526                 return;
 527
 528         if (!dbs_info->enable) {
 529                 unlock_policy_rwsem_write(cpu);
 530                 return;
 531         }
 532
 533         /* Common NORMAL_SAMPLE setup */
 534         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 535         if (!dbs_tuners_ins.powersave_bias ||
 536             sample_type == DBS_NORMAL_SAMPLE) {
 537                 dbs_check_cpu(dbs_info);
 538                 if (dbs_info->freq_lo) {
 539                         /* Setup timer for SUB_SAMPLE */
 540                         dbs_info->sample_type = DBS_SUB_SAMPLE;
 541                         delay = dbs_info->freq_hi_jiffies;
 542                 }
 543         } else {
 544                 __cpufreq_driver_target(dbs_info->cur_policy,
 545                         dbs_info->freq_lo, CPUFREQ_RELATION_H);
 546         }
 547         queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
 548         unlock_policy_rwsem_write(cpu);
 549 }
 550
 551 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
 552 {
 553         /* We want all CPUs to do sampling nearly on same jiffy */
 554         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 555         delay -= jiffies % delay;
 556
 557         dbs_info->enable = 1;
 558         ondemand_powersave_bias_init();
 559         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 560         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
 561         queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work,
 562                 delay);
 563 }
 564
 565 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
 566 {
 567         dbs_info->enable = 0;
 568         cancel_delayed_work_sync(&dbs_info->work);
 569 }
 570
 571 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 572                                    unsigned int event)
 573 {
 574         unsigned int cpu = policy->cpu;
 575         struct cpu_dbs_info_s *this_dbs_info;
 576         unsigned int j;
 577         int rc;
 578
 579         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
 580
 581         switch (event) {
 582         case CPUFREQ_GOV_START:
 583                 if ((!cpu_online(cpu)) || (!policy->cur))
 584                         return -EINVAL;
 585
 586                 if (this_dbs_info->enable) /* Already enabled */
 587                         break;
 588
 589                 mutex_lock(&dbs_mutex);
 590                 dbs_enable++;
 591
 592                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
 593                 if (rc) {
 594                         dbs_enable--;
 595                         mutex_unlock(&dbs_mutex);
 596                         return rc;
 597                 }
 598
 599                 for_each_cpu(j, policy->cpus) {
 600                         struct cpu_dbs_info_s *j_dbs_info;
 601                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
 602                         j_dbs_info->cur_policy = policy;
 603
 604                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 605                                                 &j_dbs_info->prev_cpu_wall);
 606                         if (dbs_tuners_ins.ignore_nice) {
 607                                 j_dbs_info->prev_cpu_nice =
 608                                                 kstat_cpu(j).cpustat.nice;
 609                         }
 610                 }
 611                 this_dbs_info->cpu = cpu;
 612                 /*
 613                  * Start the timerschedule work, when this governor
 614                  * is used for first time
 615                  */
 616                 if (dbs_enable == 1) {
 617                         unsigned int latency;
 618                         /* policy latency is in nS. Convert it to uS first */
 619                         latency = policy->cpuinfo.transition_latency / 1000;
 620                         if (latency == 0)
 621                                 latency = 1;
 622
 623                         def_sampling_rate =
 624                                 max(latency * LATENCY_MULTIPLIER,
 625                                     MIN_STAT_SAMPLING_RATE);
 626
 627                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
 628                 }
 629                 dbs_timer_init(this_dbs_info);
 630
 631                 mutex_unlock(&dbs_mutex);
 632                 break;
 633
 634         case CPUFREQ_GOV_STOP:
 635                 mutex_lock(&dbs_mutex);
 636                 dbs_timer_exit(this_dbs_info);
 637                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
 638                 dbs_enable--;
 639                 mutex_unlock(&dbs_mutex);
 640
 641                 break;
 642
 643         case CPUFREQ_GOV_LIMITS:
 644                 mutex_lock(&dbs_mutex);
 645                 if (policy->max < this_dbs_info->cur_policy->cur)
 646                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 647                                 policy->max, CPUFREQ_RELATION_H);
 648                 else if (policy->min > this_dbs_info->cur_policy->cur)
 649                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 650                                 policy->min, CPUFREQ_RELATION_L);
 651                 mutex_unlock(&dbs_mutex);
 652                 break;
 653         }
 654         return 0;
 655 }
 656
 657 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 658 static
 659 #endif
 660 struct cpufreq_governor cpufreq_gov_ondemand = {
 661         .name                   = "ondemand",
 662         .governor               = cpufreq_governor_dbs,
 663         .max_transition_latency = TRANSITION_LATENCY_LIMIT,
 664         .owner                  = THIS_MODULE,
 665 };
 666
 667 static int __init cpufreq_gov_dbs_init(void)
 668 {
 669         int err;
 670         cputime64_t wall;
 671         u64 idle_time;
 672         int cpu = get_cpu();
 673
 674         idle_time = get_cpu_idle_time_us(cpu, &wall);
 675         put_cpu();
 676         if (idle_time != -1ULL) {
 677                 /* Idle micro accounting is supported. Use finer thresholds */
 678                 dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
 679                 dbs_tuners_ins.down_differential =
 680                                         MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
 681         }
 682
 683         kondemand_wq = create_workqueue("kondemand");
 684         if (!kondemand_wq) {
 685                 printk(KERN_ERR "Creation of kondemand failed\n");
 686                 return -EFAULT;
 687         }
 688         err = cpufreq_register_governor(&cpufreq_gov_ondemand);
 689         if (err)
 690                 destroy_workqueue(kondemand_wq);
 691
 692         return err;
 693 }
 694
 695 static void __exit cpufreq_gov_dbs_exit(void)
 696 {
 697         cpufreq_unregister_governor(&cpufreq_gov_ondemand);
 698         destroy_workqueue(kondemand_wq);
 699 }
 700
 701
 702 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 703 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 704 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 705         "Low Latency Frequency Transition capable processors");
 706 MODULE_LICENSE("GPL");
 707
 708 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 709 fs_initcall(cpufreq_gov_dbs_init);
 710 #else
 711 module_init(cpufreq_gov_dbs_init);
 712 #endif
 713 module_exit(cpufreq_gov_dbs_exit);