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