SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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 max_load = 0;
 448         unsigned int freq_target;
 449
 450         struct cpufreq_policy *policy;
 451         unsigned int j;
 452
 453         policy = this_dbs_info->cur_policy;
 454
 455         /*
 456          * Every sampling_rate, we check, if current idle time is less
 457          * than 20% (default), then we try to increase frequency
 458          * Every sampling_rate*sampling_down_factor, we check, if current
 459          * idle time is more than 80%, then we try to decrease frequency
 460          *
 461          * Any frequency increase takes it to the maximum frequency.
 462          * Frequency reduction happens at minimum steps of
 463          * 5% (default) of maximum frequency
 464          */
 465
 466         /* Get Absolute Load */
 467         for_each_cpu(j, policy->cpus) {
 468                 struct cpu_dbs_info_s *j_dbs_info;
 469                 cputime64_t cur_wall_time, cur_idle_time;
 470                 unsigned int idle_time, wall_time;
 471
 472                 j_dbs_info = &per_cpu(cs_cpu_dbs_info, j);
 473
 474                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
 475
 476                 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
 477                                 j_dbs_info->prev_cpu_wall);
 478                 j_dbs_info->prev_cpu_wall = cur_wall_time;
 479
 480                 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
 481                                 j_dbs_info->prev_cpu_idle);
 482                 j_dbs_info->prev_cpu_idle = cur_idle_time;
 483
 484                 if (dbs_tuners_ins.ignore_nice) {
 485                         cputime64_t cur_nice;
 486                         unsigned long cur_nice_jiffies;
 487
 488                         cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
 489                                          j_dbs_info->prev_cpu_nice);
 490                         /*
 491                          * Assumption: nice time between sampling periods will
 492                          * be less than 2^32 jiffies for 32 bit sys
 493                          */
 494                         cur_nice_jiffies = (unsigned long)
 495                                         cputime64_to_jiffies64(cur_nice);
 496
 497                         j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 498                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
 499                 }
 500
 501                 if (unlikely(!wall_time || wall_time < idle_time))
 502                         continue;
 503
 504                 load = 100 * (wall_time - idle_time) / wall_time;
 505
 506                 if (load > max_load)
 507                         max_load = load;
 508         }
 509
 510         /*
 511          * break out if we 'cannot' reduce the speed as the user might
 512          * want freq_step to be zero
 513          */
 514         if (dbs_tuners_ins.freq_step == 0)
 515                 return;
 516
 517         /* Check for frequency increase */
 518         if (max_load > dbs_tuners_ins.up_threshold) {
 519                 this_dbs_info->down_skip = 0;
 520
 521                 /* if we are already at full speed then break out early */
 522                 if (this_dbs_info->requested_freq == policy->max)
 523                         return;
 524
 525                 freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
 526
 527                 /* max freq cannot be less than 100. But who knows.... */
 528                 if (unlikely(freq_target == 0))
 529                         freq_target = 5;
 530
 531                 this_dbs_info->requested_freq += freq_target;
 532                 if (this_dbs_info->requested_freq > policy->max)
 533                         this_dbs_info->requested_freq = policy->max;
 534
 535                 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
 536                         CPUFREQ_RELATION_H);
 537                 return;
 538         }
 539
 540         /*
 541          * The optimal frequency is the frequency that is the lowest that
 542          * can support the current CPU usage without triggering the up
 543          * policy. To be safe, we focus 10 points under the threshold.
 544          */
 545         if (max_load < (dbs_tuners_ins.down_threshold - 10)) {
 546                 freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100;
 547
 548                 this_dbs_info->requested_freq -= freq_target;
 549                 if (this_dbs_info->requested_freq < policy->min)
 550                         this_dbs_info->requested_freq = policy->min;
 551
 552                 /*
 553                  * if we cannot reduce the frequency anymore, break out early
 554                  */
 555                 if (policy->cur == policy->min)
 556                         return;
 557
 558                 __cpufreq_driver_target(policy, this_dbs_info->requested_freq,
 559                                 CPUFREQ_RELATION_H);
 560                 return;
 561         }
 562 }
 563
 564 static void do_dbs_timer(struct work_struct *work)
 565 {
 566         struct cpu_dbs_info_s *dbs_info =
 567                 container_of(work, struct cpu_dbs_info_s, work.work);
 568         unsigned int cpu = dbs_info->cpu;
 569
 570         /* We want all CPUs to do sampling nearly on same jiffy */
 571         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 572
 573         delay -= jiffies % delay;
 574
 575         mutex_lock(&dbs_info->timer_mutex);
 576
 577         dbs_check_cpu(dbs_info);
 578
 579         queue_delayed_work_on(cpu, kconservative_wq, &dbs_info->work, delay);
 580         mutex_unlock(&dbs_info->timer_mutex);
 581 }
 582
 583 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
 584 {
 585         /* We want all CPUs to do sampling nearly on same jiffy */
 586         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 587         delay -= jiffies % delay;
 588
 589         dbs_info->enable = 1;
 590         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
 591         queue_delayed_work_on(dbs_info->cpu, kconservative_wq, &dbs_info->work,
 592                                 delay);
 593 }
 594
 595 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
 596 {
 597         dbs_info->enable = 0;
 598         cancel_delayed_work_sync(&dbs_info->work);
 599 }
 600
 601 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 602                                    unsigned int event)
 603 {
 604         unsigned int cpu = policy->cpu;
 605         struct cpu_dbs_info_s *this_dbs_info;
 606         unsigned int j;
 607         int rc;
 608
 609         this_dbs_info = &per_cpu(cs_cpu_dbs_info, cpu);
 610
 611         switch (event) {
 612         case CPUFREQ_GOV_START:
 613                 if ((!cpu_online(cpu)) || (!policy->cur))
 614                         return -EINVAL;
 615
 616                 mutex_lock(&dbs_mutex);
 617
 618                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group_old);
 619                 if (rc) {
 620                         mutex_unlock(&dbs_mutex);
 621                         return rc;
 622                 }
 623
 624                 for_each_cpu(j, policy->cpus) {
 625                         struct cpu_dbs_info_s *j_dbs_info;
 626                         j_dbs_info = &per_cpu(cs_cpu_dbs_info, j);
 627                         j_dbs_info->cur_policy = policy;
 628
 629                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 630                                                 &j_dbs_info->prev_cpu_wall);
 631                         if (dbs_tuners_ins.ignore_nice) {
 632                                 j_dbs_info->prev_cpu_nice =
 633                                                 kstat_cpu(j).cpustat.nice;
 634                         }
 635                 }
 636                 this_dbs_info->down_skip = 0;
 637                 this_dbs_info->requested_freq = policy->cur;
 638
 639                 mutex_init(&this_dbs_info->timer_mutex);
 640                 dbs_enable++;
 641                 /*
 642                  * Start the timerschedule work, when this governor
 643                  * is used for first time
 644                  */
 645                 if (dbs_enable == 1) {
 646                         unsigned int latency;
 647                         /* policy latency is in nS. Convert it to uS first */
 648                         latency = policy->cpuinfo.transition_latency / 1000;
 649                         if (latency == 0)
 650                                 latency = 1;
 651
 652                         rc = sysfs_create_group(cpufreq_global_kobject,
 653                                                 &dbs_attr_group);
 654                         if (rc) {
 655                                 mutex_unlock(&dbs_mutex);
 656                                 return rc;
 657                         }
 658
 659                         /*
 660                          * conservative does not implement micro like ondemand
 661                          * governor, thus we are bound to jiffes/HZ
 662                          */
 663                         min_sampling_rate =
 664                                 MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
 665                         /* Bring kernel and HW constraints together */
 666                         min_sampling_rate = max(min_sampling_rate,
 667                                         MIN_LATENCY_MULTIPLIER * latency);
 668                         dbs_tuners_ins.sampling_rate =
 669                                 max(min_sampling_rate,
 670                                     latency * LATENCY_MULTIPLIER);
 671
 672                         cpufreq_register_notifier(
 673                                         &dbs_cpufreq_notifier_block,
 674                                         CPUFREQ_TRANSITION_NOTIFIER);
 675                 }
 676                 mutex_unlock(&dbs_mutex);
 677
 678                 dbs_timer_init(this_dbs_info);
 679
 680                 break;
 681
 682         case CPUFREQ_GOV_STOP:
 683                 dbs_timer_exit(this_dbs_info);
 684
 685                 mutex_lock(&dbs_mutex);
 686                 sysfs_remove_group(&policy->kobj, &dbs_attr_group_old);
 687                 dbs_enable--;
 688                 mutex_destroy(&this_dbs_info->timer_mutex);
 689
 690                 /*
 691                  * Stop the timerschedule work, when this governor
 692                  * is used for first time
 693                  */
 694                 if (dbs_enable == 0)
 695                         cpufreq_unregister_notifier(
 696                                         &dbs_cpufreq_notifier_block,
 697                                         CPUFREQ_TRANSITION_NOTIFIER);
 698
 699                 mutex_unlock(&dbs_mutex);
 700                 if (!dbs_enable)
 701                         sysfs_remove_group(cpufreq_global_kobject,
 702                                            &dbs_attr_group);
 703
 704                 break;
 705
 706         case CPUFREQ_GOV_LIMITS:
 707                 mutex_lock(&this_dbs_info->timer_mutex);
 708                 if (policy->max < this_dbs_info->cur_policy->cur)
 709                         __cpufreq_driver_target(
 710                                         this_dbs_info->cur_policy,
 711                                         policy->max, CPUFREQ_RELATION_H);
 712                 else if (policy->min > this_dbs_info->cur_policy->cur)
 713                         __cpufreq_driver_target(
 714                                         this_dbs_info->cur_policy,
 715                                         policy->min, CPUFREQ_RELATION_L);
 716                 mutex_unlock(&this_dbs_info->timer_mutex);
 717
 718                 break;
 719         }
 720         return 0;
 721 }
 722
 723 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
 724 static
 725 #endif
 726 struct cpufreq_governor cpufreq_gov_conservative = {
 727         .name                   = "conservative",
 728         .governor               = cpufreq_governor_dbs,
 729         .max_transition_latency = TRANSITION_LATENCY_LIMIT,
 730         .owner                  = THIS_MODULE,
 731 };
 732
 733 static int __init cpufreq_gov_dbs_init(void)
 734 {
 735         int err;
 736
 737         kconservative_wq = create_workqueue("kconservative");
 738         if (!kconservative_wq) {
 739                 printk(KERN_ERR "Creation of kconservative failed\n");
 740                 return -EFAULT;
 741         }
 742
 743         err = cpufreq_register_governor(&cpufreq_gov_conservative);
 744         if (err)
 745                 destroy_workqueue(kconservative_wq);
 746
 747         return err;
 748 }
 749
 750 static void __exit cpufreq_gov_dbs_exit(void)
 751 {
 752         cpufreq_unregister_governor(&cpufreq_gov_conservative);
 753         destroy_workqueue(kconservative_wq);
 754 }
 755
 756
 757 MODULE_AUTHOR("Alexander Clouter <alex@digriz.org.uk>");
 758 MODULE_DESCRIPTION("'cpufreq_conservative' - A dynamic cpufreq governor for "
 759                 "Low Latency Frequency Transition capable processors "
 760                 "optimised for use in a battery environment");
 761 MODULE_LICENSE("GPL");
 762
 763 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE
 764 fs_initcall(cpufreq_gov_dbs_init);
 765 #else
 766 module_init(cpufreq_gov_dbs_init);
 767 #endif
 768 module_exit(cpufreq_gov_dbs_exit);