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