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