drivers/cpufreq/cpufreq_conservative.c

   1 /*
   2  *  drivers/cpufreq/cpufreq_conservative.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  *            (C)  2009 Alexander Clouter <alex@digriz.org.uk>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License version 2 as
  11  * published by the Free Software Foundation.
  12  */
  13
  14 #include <linux/kernel.h>
  15 #include <linux/module.h>
  16 #include <linux/init.h>
  17 #include <linux/cpufreq.h>
  18 #include <linux/cpu.h>
  19 #include <linux/jiffies.h>
  20 #include <linux/kernel_stat.h>
  21 #include <linux/mutex.h>
  22 #include <linux/hrtimer.h>
  23 #include <linux/tick.h>
  24 #include <linux/ktime.h>
  25 #include <linux/sched.h>
  26
  27 /*
  28  * dbs is used in this file as a shortform for demandbased switching
  29  * It helps to keep variable names smaller, simpler
  30  */
  31
  32 #define DEF_FREQUENCY_UP_THRESHOLD              (80)
  33 #define DEF_FREQUENCY_DOWN_THRESHOLD            (20)
  34
  35 /*
  36  * The polling frequency of this governor depends on the capability of
  37  * the processor. Default polling frequency is 1000 times the transition
  38  * latency of the processor. The governor will work on any processor with
  39  * transition latency <= 10mS, using appropriate sampling
  40  * rate.
  41  * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
  42  * this governor will not work.
  43  * All times here are in uS.
  44  */
  45 #define MIN_SAMPLING_RATE_RATIO                 (2)
  46
  47 static unsigned int min_sampling_rate;
  48
  49 #define LATENCY_MULTIPLIER                      (1000)
  50 #define MIN_LATENCY_MULTIPLIER                  (100)
  51 #define DEF_SAMPLING_DOWN_FACTOR                (1)
  52 #define MAX_SAMPLING_DOWN_FACTOR                (10)
  53 #define TRANSITION_LATENCY_LIMIT                (10 * 1000 * 1000)
  54
  55 static void do_dbs_timer(struct work_struct *work);
  56
  57 struct cpu_dbs_info_s {
  58         cputime64_t prev_cpu_idle;
  59         cputime64_t prev_cpu_wall;
  60         cputime64_t prev_cpu_nice;
  61         struct cpufreq_policy *cur_policy;
  62         struct delayed_work work;
  63         unsigned int down_skip;
  64         unsigned int requested_freq;
  65         int cpu;
  66         unsigned int enable:1;
  67         /*
  68          * percpu mutex that serializes governor limit change with
  69          * do_dbs_timer invocation. We do not want do_dbs_timer to run
  70          * when user is changing the governor or limits.
  71          */
  72         struct mutex timer_mutex;
  73 };
  74 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cs_cpu_dbs_info);
  75
  76 static unsigned int dbs_enable; /* number of CPUs using this policy */
  77
  78 /*
  79  * dbs_mutex protects data in dbs_tuners_ins from concurrent changes on
  80  * different CPUs. It protects dbs_enable in governor start/stop.
  81  */
  82 static DEFINE_MUTEX(dbs_mutex);
  83
  84 static struct workqueue_struct  *kconservative_wq;
  85
  86 static struct dbs_tuners {
  87         unsigned int sampling_rate;
  88         unsigned int sampling_down_factor;
  89         unsigned int up_threshold;
  90         unsigned int down_threshold;
  91         unsigned int ignore_nice;
  92         unsigned int freq_step;
  93 } dbs_tuners_ins = {
  94         .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
  95         .down_threshold = DEF_FREQUENCY_DOWN_THRESHOLD,
  96         .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
  97         .ignore_nice = 0,
  98         .freq_step = 5,
  99 };
 100
 101 static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
 102                                                         cputime64_t *wall)
 103 {
 104         cputime64_t idle_time;
 105         cputime64_t cur_wall_time;
 106         cputime64_t busy_time;
 107
 108         cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
 109         busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
 110                         kstat_cpu(cpu).cpustat.system);
 111
 112         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
 113         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
 114         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
 115         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice);
 116
 117         idle_time = cputime64_sub(cur_wall_time, busy_time);
 118         if (wall)
 119                 *wall = (cputime64_t)jiffies_to_usecs(cur_wall_time);
 120
 121         return (cputime64_t)jiffies_to_usecs(idle_time);;
 122 }
 123
 124 static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
 125 {
 126         u64 idle_time = get_cpu_idle_time_us(cpu, wall);
 127
 128         if (idle_time == -1ULL)
 129                 return get_cpu_idle_time_jiffy(cpu, wall);
 130
 131         return idle_time;
 132 }
 133
 134 /* keep track of frequency transitions */
 135 static int
 136 dbs_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
 137                      void *data)
 138 {
 139         struct cpufreq_freqs *freq = data;
 140         struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cs_cpu_dbs_info,
 141                                                         freq->cpu);
 142
 143         struct cpufreq_policy *policy;
 144
 145         if (!this_dbs_info->enable)
 146                 return 0;
 147
 148         policy = this_dbs_info->cur_policy;
 149
 150         /*
 151          * we only care if our internally tracked freq moves outside
 152          * the 'valid' ranges of freqency available to us otherwise
 153          * we do not change it
 154         */
 155         if (this_dbs_info->requested_freq > policy->max
 156                         || this_dbs_info->requested_freq < policy->min)
 157                 this_dbs_info->requested_freq = freq->new;
 158
 159         return 0;
 160 }
 161
 162 static struct notifier_block dbs_cpufreq_notifier_block = {
 163         .notifier_call = dbs_cpufreq_notifier
 164 };
 165
 166 /************************** sysfs interface ************************/
 167 static ssize_t show_sampling_rate_max(struct kobject *kobj,
 168                                       struct attribute *attr, char *buf)
 169 {
 170         printk_once(KERN_INFO "CPUFREQ: conservative sampling_rate_max "
 171                     "sysfs file is deprecated - used by: %s\n", current->comm);
 172         return sprintf(buf, "%u\n", -1U);
 173 }
 174
 175 static ssize_t show_sampling_rate_min(struct kobject *kobj,
 176                                       struct attribute *attr, char *buf)
 177 {
 178         return sprintf(buf, "%u\n", min_sampling_rate);
 179 }
 180
 181 define_one_global_ro(sampling_rate_max);
 182 define_one_global_ro(sampling_rate_min);
 183
 184 /* cpufreq_conservative Governor Tunables */
 185 #define show_one(file_name, object)                                     \
 186 static ssize_t show_##file_name                                         \
 187 (struct kobject *kobj, struct attribute *attr, char *buf)               \
 188 {                                                                       \
 189         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
 190 }
 191 show_one(sampling_rate, sampling_rate);
 192 show_one(sampling_down_factor, sampling_down_factor);
 193 show_one(up_threshold, up_threshold);
 194 show_one(down_threshold, down_threshold);
 195 show_one(ignore_nice_load, ignore_nice);
 196 show_one(freq_step, freq_step);
 197
 198 /*** delete after deprecation time ***/
 199 #define DEPRECATION_MSG(file_name)                                      \
 200         printk_once(KERN_INFO "CPUFREQ: Per core conservative sysfs "   \
 201                 "interface is deprecated - " #file_name "\n");
 202
 203 #define show_one_old(file_name)                                         \
 204 static ssize_t show_##file_name##_old                                   \
 205 (struct cpufreq_policy *unused, char *buf)                              \
 206 {                                                                       \
 207         printk_once(KERN_INFO "CPUFREQ: Per core conservative sysfs "   \
 208                 "interface is deprecated - " #file_name "\n");          \
 209         return show_##file_name(NULL, NULL, buf);                       \
 210 }
 211 show_one_old(sampling_rate);
 212 show_one_old(sampling_down_factor);
 213 show_one_old(up_threshold);
 214 show_one_old(down_threshold);
 215 show_one_old(ignore_nice_load);
 216 show_one_old(freq_step);
 217 show_one_old(sampling_rate_min);
 218 show_one_old(sampling_rate_max);
 219
 220 cpufreq_freq_attr_ro_old(sampling_rate_min);
 221 cpufreq_freq_attr_ro_old(sampling_rate_max);
 222
 223 /*** delete after deprecation time ***/
 224
 225 static ssize_t store_sampling_down_factor(struct kobject *a,
 226                                           struct attribute *b,
 227                                           const char *buf, size_t count)
 228 {
 229         unsigned int input;
 230         int ret;
 231         ret = sscanf(buf, "%u", &input);
 232
 233         if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
 234                 return -EINVAL;
 235
 236         mutex_lock(&dbs_mutex);
 237         dbs_tuners_ins.sampling_down_factor = input;
 238         mutex_unlock(&dbs_mutex);
 239
 240         return count;
 241 }
 242
 243 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
 244                                    const char *buf, size_t count)
 245 {
 246         unsigned int input;
 247         int ret;
 248         ret = sscanf(buf, "%u", &input);
 249
 250         if (ret != 1)
 251                 return -EINVAL;
 252
 253         mutex_lock(&dbs_mutex);
 254         dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
 255         mutex_unlock(&dbs_mutex);
 256
 257         return count;
 258 }
 259
 260 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
 261                                   const char *buf, size_t count)
 262 {
 263         unsigned int input;
 264         int ret;
 265         ret = sscanf(buf, "%u", &input);
 266
 267         mutex_lock(&dbs_mutex);
 268         if (ret != 1 || input > 100 ||
 269                         input <= dbs_tuners_ins.down_threshold) {
 270                 mutex_unlock(&dbs_mutex);
 271                 return -EINVAL;
 272         }
 273
 274         dbs_tuners_ins.up_threshold = input;
 275         mutex_unlock(&dbs_mutex);
 276
 277         return count;
 278 }
 279
 280 static ssize_t store_down_threshold(struct kobject *a, struct attribute *b,
 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         /* cannot be lower than 11 otherwise freq will not fall */
 289         if (ret != 1 || input < 11 || input > 100 ||
 290                         input >= dbs_tuners_ins.up_threshold) {
 291                 mutex_unlock(&dbs_mutex);
 292                 return -EINVAL;
 293         }
 294
 295         dbs_tuners_ins.down_threshold = input;
 296         mutex_unlock(&dbs_mutex);
 297
 298         return count;
 299 }
 300
 301 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
 302                                       const char *buf, size_t count)
 303 {
 304         unsigned int input;
 305         int ret;
 306
 307         unsigned int j;
 308
 309         ret = sscanf(buf, "%u", &input);
 310         if (ret != 1)
 311                 return -EINVAL;
 312
 313         if (input > 1)
 314                 input = 1;
 315
 316         mutex_lock(&dbs_mutex);
 317         if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
 318                 mutex_unlock(&dbs_mutex);
 319                 return count;
 320         }
 321         dbs_tuners_ins.ignore_nice = input;
 322
 323         /* we need to re-evaluate prev_cpu_idle */
 324         for_each_online_cpu(j) {
 325                 struct cpu_dbs_info_s *dbs_info;
 326                 dbs_info = &per_cpu(cs_cpu_dbs_info, j);
 327                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 328                                                 &dbs_info->prev_cpu_wall);
 329                 if (dbs_tuners_ins.ignore_nice)
 330                         dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 331         }
 332         mutex_unlock(&dbs_mutex);
 333
 334         return count;
 335 }
 336
 337 static ssize_t store_freq_step(struct kobject *a, struct attribute *b,
 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 > 100)
 348                 input = 100;
 349
 350         /* no need to test here if freq_step is zero as the user might actually
 351          * want this, they would be crazy though :) */
 352         mutex_lock(&dbs_mutex);
 353         dbs_tuners_ins.freq_step = input;
 354         mutex_unlock(&dbs_mutex);
 355
 356         return count;
 357 }
 358
 359 define_one_global_rw(sampling_rate);
 360 define_one_global_rw(sampling_down_factor);
 361 define_one_global_rw(up_threshold);
 362 define_one_global_rw(down_threshold);
 363 define_one_global_rw(ignore_nice_load);
 364 define_one_global_rw(freq_step);
 365
 366 static struct attribute *dbs_attributes[] = {
 367         &sampling_rate_max.attr,
 368         &sampling_rate_min.attr,
 369         &sampling_rate.attr,
 370         &sampling_down_factor.attr,
 371         &up_threshold.attr,
 372         &down_threshold.attr,
 373         &ignore_nice_load.attr,
 374         &freq_step.attr,
 375         NULL
 376 };
 377
 378 static struct attribute_group dbs_attr_group = {
 379         .attrs = dbs_attributes,
 380         .name = "conservative",
 381 };
 382
 383 /*** delete after deprecation time ***/
 384
 385 #define write_one_old(file_name)                                        \
 386 static ssize_t store_##file_name##_old                                  \
 387 (struct cpufreq_policy *unused, const char *buf, size_t count)          \
 388 {                                                                       \
 389         printk_once(KERN_INFO "CPUFREQ: Per core conservative sysfs "   \
 390                 "interface is deprecated - " #file_name "\n");  \
 391         return store_##file_name(NULL, NULL, buf, count);               \
 392 }
 393 write_one_old(sampling_rate);
 394 write_one_old(sampling_down_factor);
 395 write_one_old(up_threshold);
 396 write_one_old(down_threshold);
 397 write_one_old(ignore_nice_load);
 398 write_one_old(freq_step);
 399
 400 cpufreq_freq_attr_rw_old(sampling_rate);
 401 cpufreq_freq_attr_rw_old(sampling_down_factor);
 402 cpufreq_freq_attr_rw_old(up_threshold);
 403 cpufreq_freq_attr_rw_old(down_threshold);
 404 cpufreq_freq_attr_rw_old(ignore_nice_load);
 405 cpufreq_freq_attr_rw_old(freq_step);
 406
 407 static struct attribute *dbs_attributes_old[] = {
 408         &sampling_rate_max_old.attr,
 409         &sampling_rate_min_old.attr,
 410         &sampling_rate_old.attr,
 411         &sampling_down_factor_old.attr,
 412         &up_threshold_old.attr,
 413         &down_threshold_old.attr,
 414         &ignore_nice_load_old.attr,
 415         &freq_step_old.attr,
 416         NULL
 417 };
 418
 419 static struct attribute_group dbs_attr_group_old = {
 420         .attrs = dbs_attributes_old,
 421         .name = "conservative",
 422 };
 423
 424 /*** delete after deprecation time ***/
 425
 426 /************************** sysfs end ************************/
 427
 428 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
 429 {
 430         unsigned int load = 0;
 431         unsigned int freq_target;
 432
 433         struct cpufreq_policy *policy;
 434         unsigned int j;
 435
 436         policy = this_dbs_info->cur_policy;
 437
 438         /*
 439          * Every sampling_rate, we check, if current idle time is less
 440          * than 20% (default), then we try to increase frequency
 441          * Every sampling_rate*sampling_down_factor, we check, if current
 442          * idle time is more than 80%, then we try to decrease frequency
 443          *
 444          * Any frequency increase takes it to the maximum frequency.
 445          * Frequency reduction happens at minimum steps of
 446          * 5% (default) of maximum frequency
 447          */
 448
 449         /* Get Absolute Load */
 450         for_each_cpu(j, policy->cpus) {
 451                 struct cpu_dbs_info_s *j_dbs_info;
 452                 cputime64_t cur_wall_time, cur_idle_time;
 453                 unsigned int idle_time, wall_time;
 454
 455                 j_dbs_info = &per_cpu(cs_cpu_dbs_info, j);
 456
 457                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
 458
 459                 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
 460                                 j_dbs_info->prev_cpu_wall);
 461                 j_dbs_info->prev_cpu_wall = cur_wall_time;
 462
 463                 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
 464                                 j_dbs_info->prev_cpu_idle);
 465                 j_dbs_info->prev_cpu_idle = cur_idle_time;
 466
 467                 if (dbs_tuners_ins.ignore_nice) {
 468                         cputime64_t cur_nice;
 469                         unsigned long cur_nice_jiffies;
 470
 471                         cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
 472                                          j_dbs_info->prev_cpu_nice);
 473                         /*
 474                          * Assumption: nice time between sampling periods will
 475                          * be less than 2^32 jiffies for 32 bit sys
 476                          */
 477                         cur_nice_jiffies = (unsigned long)
 478                                         cputime64_to_jiffies64(cur_nice);
 479
 480                         j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 481                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
 482                 }
 483
 484                 if (unlikely(!wall_time || wall_time < idle_time))
 485                         continue;
 486
 487                 load = 100 * (wall_time - idle_time) / wall_time;
 488         }
 489
 490         /*
 491          * break out if we 'cannot' reduce the speed as the user might
 492          * want freq_step to be zero
 493          */
 494         if (dbs_tuners_ins.freq_step == 0)
 495                 return;
 496
 497         /* Check for frequency increase */
 498         if (load > dbs_tuners_ins.up_threshold) {
 499                 this_dbs_info->down_skip = 0;
 500
 501                 /* if we are already at full speed then break out early */
 502                 if (this_dbs_info->requested_freq == policy->max)
 503                         return;
 504
 505                 freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
 506
 507                 /* max freq cannot be less than 100. But who knows.... */
 508                 if (unlikely(freq_target == 0))
 509                         freq_target = 5;
 510
 511                 this_dbs_info->requested_freq += freq_target;
 512                 if (this_dbs_info->requested_freq > policy->max)
 513                         this_dbs_info->requested_freq = policy->max;
 514
 515                 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
 516                         CPUFREQ_RELATION_H);
 517                 return;
 518         }
 519
 520         /*
 521          * The optimal frequency is the frequency that is the lowest that
 522          * can support the current CPU usage without triggering the up
 523          * policy. To be safe, we focus 10 points under the threshold.
 524          */
 525         if (load < (dbs_tuners_ins.down_threshold - 10)) {
 526                 freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
 527
 528                 this_dbs_info->requested_freq -= freq_target;
 529                 if (this_dbs_info->requested_freq < policy->min)
 530                         this_dbs_info->requested_freq = policy->min;
 531
 532                 /*
 533                  * if we cannot reduce the frequency anymore, break out early
 534                  */
 535                 if (policy->cur == policy->min)
 536                         return;
 537
 538                 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
 539                                 CPUFREQ_RELATION_H);
 540                 return;
 541         }
 542 }
 543
 544 static void do_dbs_timer(struct work_struct *work)
 545 {
 546         struct cpu_dbs_info_s *dbs_info =
 547                 container_of(work, struct cpu_dbs_info_s, work.work);
 548         unsigned int cpu = dbs_info->cpu;
 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
 553         delay -= jiffies % delay;
 554
 555         mutex_lock(&dbs_info->timer_mutex);
 556
 557         dbs_check_cpu(dbs_info);
 558
 559         queue_delayed_work_on(cpu, kconservative_wq, &dbs_info->work, delay);
 560         mutex_unlock(&dbs_info->timer_mutex);
 561 }
 562
 563 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
 564 {
 565         /* We want all CPUs to do sampling nearly on same jiffy */
 566         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 567         delay -= jiffies % delay;
 568
 569         dbs_info->enable = 1;
 570         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
 571         queue_delayed_work_on(dbs_info->cpu, kconservative_wq, &dbs_info->work,
 572                                 delay);
 573 }
 574
 575 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
 576 {
 577         dbs_info->enable = 0;
 578         cancel_delayed_work_sync(&dbs_info->work);
 579 }
 580
 581 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 582                                    unsigned int event)
 583 {
 584         unsigned int cpu = policy->cpu;
 585         struct cpu_dbs_info_s *this_dbs_info;
 586         unsigned int j;
 587         int rc;
 588
 589         this_dbs_info = &per_cpu(cs_cpu_dbs_info, cpu);
 590
 591         switch (event) {
 592         case CPUFREQ_GOV_START:
 593                 if ((!cpu_online(cpu)) || (!policy->cur))
 594                         return -EINVAL;
 595
 596                 mutex_lock(&dbs_mutex);
 597
 598                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group_old);
 599                 if (rc) {
 600                         mutex_unlock(&dbs_mutex);
 601                         return rc;
 602                 }
 603
 604                 for_each_cpu(j, policy->cpus) {
 605                         struct cpu_dbs_info_s *j_dbs_info;
 606                         j_dbs_info = &per_cpu(cs_cpu_dbs_info, j);
 607                         j_dbs_info->cur_policy = policy;
 608
 609                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 610                                                 &j_dbs_info->prev_cpu_wall);
 611                         if (dbs_tuners_ins.ignore_nice) {
 612                                 j_dbs_info->prev_cpu_nice =
 613                                                 kstat_cpu(j).cpustat.nice;
 614                         }
 615                 }
 616                 this_dbs_info->down_skip = 0;
 617                 this_dbs_info->requested_freq = policy->cur;
 618
 619                 mutex_init(&this_dbs_info->timer_mutex);
 620                 dbs_enable++;
 621                 /*
 622                  * Start the timerschedule work, when this governor
 623                  * is used for first time
 624                  */
 625                 if (dbs_enable == 1) {
 626                         unsigned int latency;
 627                         /* policy latency is in nS. Convert it to uS first */
 628                         latency = policy->cpuinfo.transition_latency / 1000;
 629                         if (latency == 0)
 630                                 latency = 1;
 631
 632                         rc = sysfs_create_group(cpufreq_global_kobject,
 633                                                 &dbs_attr_group);
 634                         if (rc) {
 635                                 mutex_unlock(&dbs_mutex);
 636                                 return rc;
 637                         }
 638
 639                         /*
 640                          * conservative does not implement micro like ondemand
 641                          * governor, thus we are bound to jiffes/HZ
 642                          */
 643                         min_sampling_rate =
 644                                 MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
 645                         /* Bring kernel and HW constraints together */
 646                         min_sampling_rate = max(min_sampling_rate,
 647                                         MIN_LATENCY_MULTIPLIER * latency);
 648                         dbs_tuners_ins.sampling_rate =
 649                                 max(min_sampling_rate,
 650                                     latency * LATENCY_MULTIPLIER);
 651
 652                         cpufreq_register_notifier(
 653                                         &dbs_cpufreq_notifier_block,
 654                                         CPUFREQ_TRANSITION_NOTIFIER);
 655                 }
 656                 mutex_unlock(&dbs_mutex);
 657
 658                 dbs_timer_init(this_dbs_info);
 659
 660                 break;
 661
 662         case CPUFREQ_GOV_STOP:
 663                 dbs_timer_exit(this_dbs_info);
 664
 665                 mutex_lock(&dbs_mutex);
 666                 sysfs_remove_group(&policy->kobj, &dbs_attr_group_old);
 667                 dbs_enable--;
 668                 mutex_destroy(&this_dbs_info->timer_mutex);
 669
 670                 /*
 671                  * Stop the timerschedule work, when this governor
 672                  * is used for first time
 673                  */
 674                 if (dbs_enable == 0)
 675                         cpufreq_unregister_notifier(
 676                                         &dbs_cpufreq_notifier_block,
 677                                         CPUFREQ_TRANSITION_NOTIFIER);
 678
 679                 mutex_unlock(&dbs_mutex);
 680                 if (!dbs_enable)
 681                         sysfs_remove_group(cpufreq_global_kobject,
 682                                            &dbs_attr_group);
 683
 684                 break;
 685
 686         case CPUFREQ_GOV_LIMITS:
 687                 mutex_lock(&this_dbs_info->timer_mutex);
 688                 if (policy->max < this_dbs_info->cur_policy->cur)
 689                         __cpufreq_driver_target(
 690                                         this_dbs_info->cur_policy,
 691                                         policy->max, CPUFREQ_RELATION_H);
 692                 else if (policy->min > this_dbs_info->cur_policy->cur)
 693                         __cpufreq_driver_target(
 694                                         this_dbs_info->cur_policy,
 695                                         policy->min, CPUFREQ_RELATION_L);
 696                 mutex_unlock(&this_dbs_info->timer_mutex);
 697
 698                 break;
 699         }
 700         return 0;
 701 }
 702
 703 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
 704 static
 705 #endif
 706 struct cpufreq_governor cpufreq_gov_conservative = {
 707         .name                   = "conservative",
 708         .governor               = cpufreq_governor_dbs,
 709         .max_transition_latency = TRANSITION_LATENCY_LIMIT,
 710         .owner                  = THIS_MODULE,
 711 };
 712
 713 static int __init cpufreq_gov_dbs_init(void)
 714 {
 715         int err;
 716
 717         kconservative_wq = create_workqueue("kconservative");
 718         if (!kconservative_wq) {
 719                 printk(KERN_ERR "Creation of kconservative failed\n");
 720                 return -EFAULT;
 721         }
 722
 723         err = cpufreq_register_governor(&cpufreq_gov_conservative);
 724         if (err)
 725                 destroy_workqueue(kconservative_wq);
 726
 727         return err;
 728 }
 729
 730 static void __exit cpufreq_gov_dbs_exit(void)
 731 {
 732         cpufreq_unregister_governor(&cpufreq_gov_conservative);
 733         destroy_workqueue(kconservative_wq);
 734 }
 735
 736
 737 MODULE_AUTHOR("Alexander Clouter <alex@digriz.org.uk>");
 738 MODULE_DESCRIPTION("'cpufreq_conservative' - A dynamic cpufreq governor for "
 739                 "Low Latency Frequency Transition capable processors "
 740                 "optimised for use in a battery environment");
 741 MODULE_LICENSE("GPL");
 742
 743 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
 744 fs_initcall(cpufreq_gov_dbs_init);
 745 #else
 746 module_init(cpufreq_gov_dbs_init);
 747 #endif
 748 module_exit(cpufreq_gov_dbs_exit);