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 define_one_ro(_name)            \
 182 static struct global_attr _name =       \
 183 __ATTR(_name, 0444, show_##_name, NULL)
 184
 185 define_one_ro(sampling_rate_max);
 186 define_one_ro(sampling_rate_min);
 187
 188 /* cpufreq_conservative Governor Tunables */
 189 #define show_one(file_name, object)                                     \
 190 static ssize_t show_##file_name                                         \
 191 (struct kobject *kobj, struct attribute *attr, char *buf)               \
 192 {                                                                       \
 193         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
 194 }
 195 show_one(sampling_rate, sampling_rate);
 196 show_one(sampling_down_factor, sampling_down_factor);
 197 show_one(up_threshold, up_threshold);
 198 show_one(down_threshold, down_threshold);
 199 show_one(ignore_nice_load, ignore_nice);
 200 show_one(freq_step, freq_step);
 201
 202 /*** delete after deprecation time ***/
 203 #define DEPRECATION_MSG(file_name)                                      \
 204         printk_once(KERN_INFO "CPUFREQ: Per core conservative sysfs "   \
 205                 "interface is deprecated - " #file_name "\n");
 206
 207 #define show_one_old(file_name)                                         \
 208 static ssize_t show_##file_name##_old                                   \
 209 (struct cpufreq_policy *unused, char *buf)                              \
 210 {                                                                       \
 211         printk_once(KERN_INFO "CPUFREQ: Per core conservative sysfs "   \
 212                 "interface is deprecated - " #file_name "\n");          \
 213         return show_##file_name(NULL, NULL, buf);                       \
 214 }
 215 show_one_old(sampling_rate);
 216 show_one_old(sampling_down_factor);
 217 show_one_old(up_threshold);
 218 show_one_old(down_threshold);
 219 show_one_old(ignore_nice_load);
 220 show_one_old(freq_step);
 221 show_one_old(sampling_rate_min);
 222 show_one_old(sampling_rate_max);
 223
 224 #define define_one_ro_old(object, _name)        \
 225 static struct freq_attr object =                \
 226 __ATTR(_name, 0444, show_##_name##_old, NULL)
 227
 228 define_one_ro_old(sampling_rate_min_old, sampling_rate_min);
 229 define_one_ro_old(sampling_rate_max_old, sampling_rate_max);
 230
 231 /*** delete after deprecation time ***/
 232
 233 static ssize_t store_sampling_down_factor(struct kobject *a,
 234                                           struct attribute *b,
 235                                           const char *buf, size_t count)
 236 {
 237         unsigned int input;
 238         int ret;
 239         ret = sscanf(buf, "%u", &input);
 240
 241         if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
 242                 return -EINVAL;
 243
 244         mutex_lock(&dbs_mutex);
 245         dbs_tuners_ins.sampling_down_factor = input;
 246         mutex_unlock(&dbs_mutex);
 247
 248         return count;
 249 }
 250
 251 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
 252                                    const char *buf, size_t count)
 253 {
 254         unsigned int input;
 255         int ret;
 256         ret = sscanf(buf, "%u", &input);
 257
 258         if (ret != 1)
 259                 return -EINVAL;
 260
 261         mutex_lock(&dbs_mutex);
 262         dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
 263         mutex_unlock(&dbs_mutex);
 264
 265         return count;
 266 }
 267
 268 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
 269                                   const char *buf, size_t count)
 270 {
 271         unsigned int input;
 272         int ret;
 273         ret = sscanf(buf, "%u", &input);
 274
 275         mutex_lock(&dbs_mutex);
 276         if (ret != 1 || input > 100 ||
 277                         input <= dbs_tuners_ins.down_threshold) {
 278                 mutex_unlock(&dbs_mutex);
 279                 return -EINVAL;
 280         }
 281
 282         dbs_tuners_ins.up_threshold = input;
 283         mutex_unlock(&dbs_mutex);
 284
 285         return count;
 286 }
 287
 288 static ssize_t store_down_threshold(struct kobject *a, struct attribute *b,
 289                                     const char *buf, size_t count)
 290 {
 291         unsigned int input;
 292         int ret;
 293         ret = sscanf(buf, "%u", &input);
 294
 295         mutex_lock(&dbs_mutex);
 296         /* cannot be lower than 11 otherwise freq will not fall */
 297         if (ret != 1 || input < 11 || input > 100 ||
 298                         input >= dbs_tuners_ins.up_threshold) {
 299                 mutex_unlock(&dbs_mutex);
 300                 return -EINVAL;
 301         }
 302
 303         dbs_tuners_ins.down_threshold = input;
 304         mutex_unlock(&dbs_mutex);
 305
 306         return count;
 307 }
 308
 309 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
 310                                       const char *buf, size_t count)
 311 {
 312         unsigned int input;
 313         int ret;
 314
 315         unsigned int j;
 316
 317         ret = sscanf(buf, "%u", &input);
 318         if (ret != 1)
 319                 return -EINVAL;
 320
 321         if (input > 1)
 322                 input = 1;
 323
 324         mutex_lock(&dbs_mutex);
 325         if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
 326                 mutex_unlock(&dbs_mutex);
 327                 return count;
 328         }
 329         dbs_tuners_ins.ignore_nice = input;
 330
 331         /* we need to re-evaluate prev_cpu_idle */
 332         for_each_online_cpu(j) {
 333                 struct cpu_dbs_info_s *dbs_info;
 334                 dbs_info = &per_cpu(cs_cpu_dbs_info, j);
 335                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 336                                                 &dbs_info->prev_cpu_wall);
 337                 if (dbs_tuners_ins.ignore_nice)
 338                         dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 339         }
 340         mutex_unlock(&dbs_mutex);
 341
 342         return count;
 343 }
 344
 345 static ssize_t store_freq_step(struct kobject *a, struct attribute *b,
 346                                const char *buf, size_t count)
 347 {
 348         unsigned int input;
 349         int ret;
 350         ret = sscanf(buf, "%u", &input);
 351
 352         if (ret != 1)
 353                 return -EINVAL;
 354
 355         if (input > 100)
 356                 input = 100;
 357
 358         /* no need to test here if freq_step is zero as the user might actually
 359          * want this, they would be crazy though :) */
 360         mutex_lock(&dbs_mutex);
 361         dbs_tuners_ins.freq_step = input;
 362         mutex_unlock(&dbs_mutex);
 363
 364         return count;
 365 }
 366
 367 #define define_one_rw(_name) \
 368 static struct global_attr _name = \
 369 __ATTR(_name, 0644, show_##_name, store_##_name)
 370
 371 define_one_rw(sampling_rate);
 372 define_one_rw(sampling_down_factor);
 373 define_one_rw(up_threshold);
 374 define_one_rw(down_threshold);
 375 define_one_rw(ignore_nice_load);
 376 define_one_rw(freq_step);
 377
 378 static struct attribute *dbs_attributes[] = {
 379         &sampling_rate_max.attr,
 380         &sampling_rate_min.attr,
 381         &sampling_rate.attr,
 382         &sampling_down_factor.attr,
 383         &up_threshold.attr,
 384         &down_threshold.attr,
 385         &ignore_nice_load.attr,
 386         &freq_step.attr,
 387         NULL
 388 };
 389
 390 static struct attribute_group dbs_attr_group = {
 391         .attrs = dbs_attributes,
 392         .name = "conservative",
 393 };
 394
 395 /*** delete after deprecation time ***/
 396
 397 #define write_one_old(file_name)                                        \
 398 static ssize_t store_##file_name##_old                                  \
 399 (struct cpufreq_policy *unused, const char *buf, size_t count)          \
 400 {                                                                       \
 401         printk_once(KERN_INFO "CPUFREQ: Per core conservative sysfs "   \
 402                 "interface is deprecated - " #file_name "\n");  \
 403         return store_##file_name(NULL, NULL, buf, count);               \
 404 }
 405 write_one_old(sampling_rate);
 406 write_one_old(sampling_down_factor);
 407 write_one_old(up_threshold);
 408 write_one_old(down_threshold);
 409 write_one_old(ignore_nice_load);
 410 write_one_old(freq_step);
 411
 412 #define define_one_rw_old(object, _name)        \
 413 static struct freq_attr object =                \
 414 __ATTR(_name, 0644, show_##_name##_old, store_##_name##_old)
 415
 416 define_one_rw_old(sampling_rate_old, sampling_rate);
 417 define_one_rw_old(sampling_down_factor_old, sampling_down_factor);
 418 define_one_rw_old(up_threshold_old, up_threshold);
 419 define_one_rw_old(down_threshold_old, down_threshold);
 420 define_one_rw_old(ignore_nice_load_old, ignore_nice_load);
 421 define_one_rw_old(freq_step_old, freq_step);
 422
 423 static struct attribute *dbs_attributes_old[] = {
 424         &sampling_rate_max_old.attr,
 425         &sampling_rate_min_old.attr,
 426         &sampling_rate_old.attr,
 427         &sampling_down_factor_old.attr,
 428         &up_threshold_old.attr,
 429         &down_threshold_old.attr,
 430         &ignore_nice_load_old.attr,
 431         &freq_step_old.attr,
 432         NULL
 433 };
 434
 435 static struct attribute_group dbs_attr_group_old = {
 436         .attrs = dbs_attributes_old,
 437         .name = "conservative",
 438 };
 439
 440 /*** delete after deprecation time ***/
 441
 442 /************************** sysfs end ************************/
 443
 444 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
 445 {
 446         unsigned int load = 0;
 447         unsigned int freq_target;
 448
 449         struct cpufreq_policy *policy;
 450         unsigned int j;
 451
 452         policy = this_dbs_info->cur_policy;
 453
 454         /*
 455          * Every sampling_rate, we check, if current idle time is less
 456          * than 20% (default), then we try to increase frequency
 457          * Every sampling_rate*sampling_down_factor, we check, if current
 458          * idle time is more than 80%, then we try to decrease frequency
 459          *
 460          * Any frequency increase takes it to the maximum frequency.
 461          * Frequency reduction happens at minimum steps of
 462          * 5% (default) of maximum frequency
 463          */
 464
 465         /* Get Absolute Load */
 466         for_each_cpu(j, policy->cpus) {
 467                 struct cpu_dbs_info_s *j_dbs_info;
 468                 cputime64_t cur_wall_time, cur_idle_time;
 469                 unsigned int idle_time, wall_time;
 470
 471                 j_dbs_info = &per_cpu(cs_cpu_dbs_info, j);
 472
 473                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
 474
 475                 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
 476                                 j_dbs_info->prev_cpu_wall);
 477                 j_dbs_info->prev_cpu_wall = cur_wall_time;
 478
 479                 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
 480                                 j_dbs_info->prev_cpu_idle);
 481                 j_dbs_info->prev_cpu_idle = cur_idle_time;
 482
 483                 if (dbs_tuners_ins.ignore_nice) {
 484                         cputime64_t cur_nice;
 485                         unsigned long cur_nice_jiffies;
 486
 487                         cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
 488                                          j_dbs_info->prev_cpu_nice);
 489                         /*
 490                          * Assumption: nice time between sampling periods will
 491                          * be less than 2^32 jiffies for 32 bit sys
 492                          */
 493                         cur_nice_jiffies = (unsigned long)
 494                                         cputime64_to_jiffies64(cur_nice);
 495
 496                         j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 497                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
 498                 }
 499
 500                 if (unlikely(!wall_time || wall_time < idle_time))
 501                         continue;
 502
 503                 load = 100 * (wall_time - idle_time) / wall_time;
 504         }
 505
 506         /*
 507          * break out if we 'cannot' reduce the speed as the user might
 508          * want freq_step to be zero
 509          */
 510         if (dbs_tuners_ins.freq_step == 0)
 511                 return;
 512
 513         /* Check for frequency increase */
 514         if (load > dbs_tuners_ins.up_threshold) {
 515                 this_dbs_info->down_skip = 0;
 516
 517                 /* if we are already at full speed then break out early */
 518                 if (this_dbs_info->requested_freq == policy->max)
 519                         return;
 520
 521                 freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
 522
 523                 /* max freq cannot be less than 100. But who knows.... */
 524                 if (unlikely(freq_target == 0))
 525                         freq_target = 5;
 526
 527                 this_dbs_info->requested_freq += freq_target;
 528                 if (this_dbs_info->requested_freq > policy->max)
 529                         this_dbs_info->requested_freq = policy->max;
 530
 531                 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
 532                         CPUFREQ_RELATION_H);
 533                 return;
 534         }
 535
 536         /*
 537          * The optimal frequency is the frequency that is the lowest that
 538          * can support the current CPU usage without triggering the up
 539          * policy. To be safe, we focus 10 points under the threshold.
 540          */
 541         if (load < (dbs_tuners_ins.down_threshold - 10)) {
 542                 freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
 543
 544                 this_dbs_info->requested_freq -= freq_target;
 545                 if (this_dbs_info->requested_freq < policy->min)
 546                         this_dbs_info->requested_freq = policy->min;
 547
 548                 /*
 549                  * if we cannot reduce the frequency anymore, break out early
 550                  */
 551                 if (policy->cur == policy->min)
 552                         return;
 553
 554                 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
 555                                 CPUFREQ_RELATION_H);
 556                 return;
 557         }
 558 }
 559
 560 static void do_dbs_timer(struct work_struct *work)
 561 {
 562         struct cpu_dbs_info_s *dbs_info =
 563                 container_of(work, struct cpu_dbs_info_s, work.work);
 564         unsigned int cpu = dbs_info->cpu;
 565
 566         /* We want all CPUs to do sampling nearly on same jiffy */
 567         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 568
 569         delay -= jiffies % delay;
 570
 571         mutex_lock(&dbs_info->timer_mutex);
 572
 573         dbs_check_cpu(dbs_info);
 574
 575         queue_delayed_work_on(cpu, kconservative_wq, &dbs_info->work, delay);
 576         mutex_unlock(&dbs_info->timer_mutex);
 577 }
 578
 579 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
 580 {
 581         /* We want all CPUs to do sampling nearly on same jiffy */
 582         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 583         delay -= jiffies % delay;
 584
 585         dbs_info->enable = 1;
 586         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
 587         queue_delayed_work_on(dbs_info->cpu, kconservative_wq, &dbs_info->work,
 588                                 delay);
 589 }
 590
 591 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
 592 {
 593         dbs_info->enable = 0;
 594         cancel_delayed_work_sync(&dbs_info->work);
 595 }
 596
 597 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 598                                    unsigned int event)
 599 {
 600         unsigned int cpu = policy->cpu;
 601         struct cpu_dbs_info_s *this_dbs_info;
 602         unsigned int j;
 603         int rc;
 604
 605         this_dbs_info = &per_cpu(cs_cpu_dbs_info, cpu);
 606
 607         switch (event) {
 608         case CPUFREQ_GOV_START:
 609                 if ((!cpu_online(cpu)) || (!policy->cur))
 610                         return -EINVAL;
 611
 612                 mutex_lock(&dbs_mutex);
 613
 614                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group_old);
 615                 if (rc) {
 616                         mutex_unlock(&dbs_mutex);
 617                         return rc;
 618                 }
 619
 620                 for_each_cpu(j, policy->cpus) {
 621                         struct cpu_dbs_info_s *j_dbs_info;
 622                         j_dbs_info = &per_cpu(cs_cpu_dbs_info, j);
 623                         j_dbs_info->cur_policy = policy;
 624
 625                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 626                                                 &j_dbs_info->prev_cpu_wall);
 627                         if (dbs_tuners_ins.ignore_nice) {
 628                                 j_dbs_info->prev_cpu_nice =
 629                                                 kstat_cpu(j).cpustat.nice;
 630                         }
 631                 }
 632                 this_dbs_info->down_skip = 0;
 633                 this_dbs_info->requested_freq = policy->cur;
 634
 635                 mutex_init(&this_dbs_info->timer_mutex);
 636                 dbs_enable++;
 637                 /*
 638                  * Start the timerschedule work, when this governor
 639                  * is used for first time
 640                  */
 641                 if (dbs_enable == 1) {
 642                         unsigned int latency;
 643                         /* policy latency is in nS. Convert it to uS first */
 644                         latency = policy->cpuinfo.transition_latency / 1000;
 645                         if (latency == 0)
 646                                 latency = 1;
 647
 648                         rc = sysfs_create_group(cpufreq_global_kobject,
 649                                                 &dbs_attr_group);
 650                         if (rc) {
 651                                 mutex_unlock(&dbs_mutex);
 652                                 return rc;
 653                         }
 654
 655                         /*
 656                          * conservative does not implement micro like ondemand
 657                          * governor, thus we are bound to jiffes/HZ
 658                          */
 659                         min_sampling_rate =
 660                                 MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
 661                         /* Bring kernel and HW constraints together */
 662                         min_sampling_rate = max(min_sampling_rate,
 663                                         MIN_LATENCY_MULTIPLIER * latency);
 664                         dbs_tuners_ins.sampling_rate =
 665                                 max(min_sampling_rate,
 666                                     latency * LATENCY_MULTIPLIER);
 667
 668                         cpufreq_register_notifier(
 669                                         &dbs_cpufreq_notifier_block,
 670                                         CPUFREQ_TRANSITION_NOTIFIER);
 671                 }
 672                 mutex_unlock(&dbs_mutex);
 673
 674                 dbs_timer_init(this_dbs_info);
 675
 676                 break;
 677
 678         case CPUFREQ_GOV_STOP:
 679                 dbs_timer_exit(this_dbs_info);
 680
 681                 mutex_lock(&dbs_mutex);
 682                 sysfs_remove_group(&policy->kobj, &dbs_attr_group_old);
 683                 dbs_enable--;
 684                 mutex_destroy(&this_dbs_info->timer_mutex);
 685
 686                 /*
 687                  * Stop the timerschedule work, when this governor
 688                  * is used for first time
 689                  */
 690                 if (dbs_enable == 0)
 691                         cpufreq_unregister_notifier(
 692                                         &dbs_cpufreq_notifier_block,
 693                                         CPUFREQ_TRANSITION_NOTIFIER);
 694
 695                 mutex_unlock(&dbs_mutex);
 696                 if (!dbs_enable)
 697                         sysfs_remove_group(cpufreq_global_kobject,
 698                                            &dbs_attr_group);
 699
 700                 break;
 701
 702         case CPUFREQ_GOV_LIMITS:
 703                 mutex_lock(&this_dbs_info->timer_mutex);
 704                 if (policy->max < this_dbs_info->cur_policy->cur)
 705                         __cpufreq_driver_target(
 706                                         this_dbs_info->cur_policy,
 707                                         policy->max, CPUFREQ_RELATION_H);
 708                 else if (policy->min > this_dbs_info->cur_policy->cur)
 709                         __cpufreq_driver_target(
 710                                         this_dbs_info->cur_policy,
 711                                         policy->min, CPUFREQ_RELATION_L);
 712                 mutex_unlock(&this_dbs_info->timer_mutex);
 713
 714                 break;
 715         }
 716         return 0;
 717 }
 718
 719 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
 720 static
 721 #endif
 722 struct cpufreq_governor cpufreq_gov_conservative = {
 723         .name                   = "conservative",
 724         .governor               = cpufreq_governor_dbs,
 725         .max_transition_latency = TRANSITION_LATENCY_LIMIT,
 726         .owner                  = THIS_MODULE,
 727 };
 728
 729 static int __init cpufreq_gov_dbs_init(void)
 730 {
 731         int err;
 732
 733         kconservative_wq = create_workqueue("kconservative");
 734         if (!kconservative_wq) {
 735                 printk(KERN_ERR "Creation of kconservative failed\n");
 736                 return -EFAULT;
 737         }
 738
 739         err = cpufreq_register_governor(&cpufreq_gov_conservative);
 740         if (err)
 741                 destroy_workqueue(kconservative_wq);
 742
 743         return err;
 744 }
 745
 746 static void __exit cpufreq_gov_dbs_exit(void)
 747 {
 748         cpufreq_unregister_governor(&cpufreq_gov_conservative);
 749         destroy_workqueue(kconservative_wq);
 750 }
 751
 752
 753 MODULE_AUTHOR("Alexander Clouter <alex@digriz.org.uk>");
 754 MODULE_DESCRIPTION("'cpufreq_conservative' - A dynamic cpufreq governor for "
 755                 "Low Latency Frequency Transition capable processors "
 756                 "optimised for use in a battery environment");
 757 MODULE_LICENSE("GPL");
 758
 759 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
 760 fs_initcall(cpufreq_gov_dbs_init);
 761 #else
 762 module_init(cpufreq_gov_dbs_init);
 763 #endif
 764 module_exit(cpufreq_gov_dbs_exit);