drivers/cpufreq/cpufreq_ondemand.c

   1 /*
   2  *  drivers/cpufreq/cpufreq_ondemand.c
   3  *
   4  *  Copyright (C)  2001 Russell King
   5  *            (C)  2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
   6  *                      Jun Nakajima <jun.nakajima@intel.com>
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 as
  10  * published by the Free Software Foundation.
  11  */
  12
  13 #include <linux/kernel.h>
  14 #include <linux/module.h>
  15 #include <linux/init.h>
  16 #include <linux/cpufreq.h>
  17 #include <linux/cpu.h>
  18 #include <linux/jiffies.h>
  19 #include <linux/kernel_stat.h>
  20 #include <linux/mutex.h>
  21 #include <linux/hrtimer.h>
  22 #include <linux/tick.h>
  23 #include <linux/ktime.h>
  24 #include <linux/sched.h>
  25
  26 /*
  27  * dbs is used in this file as a shortform for demandbased switching
  28  * It helps to keep variable names smaller, simpler
  29  */
  30
  31 #define DEF_FREQUENCY_DOWN_DIFFERENTIAL         (10)
  32 #define DEF_FREQUENCY_UP_THRESHOLD              (80)
  33 #define MICRO_FREQUENCY_DOWN_DIFFERENTIAL       (3)
  34 #define MICRO_FREQUENCY_UP_THRESHOLD            (95)
  35 #define MICRO_FREQUENCY_MIN_SAMPLE_RATE         (10000)
  36 #define MIN_FREQUENCY_UP_THRESHOLD              (11)
  37 #define MAX_FREQUENCY_UP_THRESHOLD              (100)
  38
  39 /*
  40  * The polling frequency of this governor depends on the capability of
  41  * the processor. Default polling frequency is 1000 times the transition
  42  * latency of the processor. The governor will work on any processor with
  43  * transition latency <= 10mS, using appropriate sampling
  44  * rate.
  45  * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
  46  * this governor will not work.
  47  * All times here are in uS.
  48  */
  49 #define MIN_SAMPLING_RATE_RATIO                 (2)
  50
  51 static unsigned int min_sampling_rate;
  52
  53 #define LATENCY_MULTIPLIER                      (1000)
  54 #define MIN_LATENCY_MULTIPLIER                  (100)
  55 #define TRANSITION_LATENCY_LIMIT                (10 * 1000 * 1000)
  56
  57 static void do_dbs_timer(struct work_struct *work);
  58 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
  59                                 unsigned int event);
  60
  61 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
  62 static
  63 #endif
  64 struct cpufreq_governor cpufreq_gov_ondemand = {
  65        .name                   = "ondemand",
  66        .governor               = cpufreq_governor_dbs,
  67        .max_transition_latency = TRANSITION_LATENCY_LIMIT,
  68        .owner                  = THIS_MODULE,
  69 };
  70
  71 /* Sampling types */
  72 enum {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE};
  73
  74 struct cpu_dbs_info_s {
  75         cputime64_t prev_cpu_idle;
  76         cputime64_t prev_cpu_wall;
  77         cputime64_t prev_cpu_nice;
  78         struct cpufreq_policy *cur_policy;
  79         struct delayed_work work;
  80         struct cpufreq_frequency_table *freq_table;
  81         unsigned int freq_lo;
  82         unsigned int freq_lo_jiffies;
  83         unsigned int freq_hi_jiffies;
  84         int cpu;
  85         unsigned int sample_type:1;
  86         /*
  87          * percpu mutex that serializes governor limit change with
  88          * do_dbs_timer invocation. We do not want do_dbs_timer to run
  89          * when user is changing the governor or limits.
  90          */
  91         struct mutex timer_mutex;
  92 };
  93 static DEFINE_PER_CPU(struct cpu_dbs_info_s, od_cpu_dbs_info);
  94
  95 static unsigned int dbs_enable; /* number of CPUs using this policy */
  96
  97 /*
  98  * dbs_mutex protects data in dbs_tuners_ins from concurrent changes on
  99  * different CPUs. It protects dbs_enable in governor start/stop.
 100  */
 101 static DEFINE_MUTEX(dbs_mutex);
 102
 103 static struct workqueue_struct  *kondemand_wq;
 104
 105 static struct dbs_tuners {
 106         unsigned int sampling_rate;
 107         unsigned int up_threshold;
 108         unsigned int down_differential;
 109         unsigned int ignore_nice;
 110         unsigned int powersave_bias;
 111 } dbs_tuners_ins = {
 112         .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
 113         .down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL,
 114         .ignore_nice = 0,
 115         .powersave_bias = 0,
 116 };
 117
 118 static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
 119                                                         cputime64_t *wall)
 120 {
 121         cputime64_t idle_time;
 122         cputime64_t cur_wall_time;
 123         cputime64_t busy_time;
 124
 125         cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
 126         busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
 127                         kstat_cpu(cpu).cpustat.system);
 128
 129         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
 130         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
 131         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
 132         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice);
 133
 134         idle_time = cputime64_sub(cur_wall_time, busy_time);
 135         if (wall)
 136                 *wall = (cputime64_t)jiffies_to_usecs(cur_wall_time);
 137
 138         return (cputime64_t)jiffies_to_usecs(idle_time);
 139 }
 140
 141 static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
 142 {
 143         u64 idle_time = get_cpu_idle_time_us(cpu, wall);
 144
 145         if (idle_time == -1ULL)
 146                 return get_cpu_idle_time_jiffy(cpu, wall);
 147
 148         return idle_time;
 149 }
 150
 151 /*
 152  * Find right freq to be set now with powersave_bias on.
 153  * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
 154  * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
 155  */
 156 static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
 157                                           unsigned int freq_next,
 158                                           unsigned int relation)
 159 {
 160         unsigned int freq_req, freq_reduc, freq_avg;
 161         unsigned int freq_hi, freq_lo;
 162         unsigned int index = 0;
 163         unsigned int jiffies_total, jiffies_hi, jiffies_lo;
 164         struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
 165                                                    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_cpu(int cpu)
 207 {
 208         struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 209         dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
 210         dbs_info->freq_lo = 0;
 211 }
 212
 213 static void ondemand_powersave_bias_init(void)
 214 {
 215         int i;
 216         for_each_online_cpu(i) {
 217                 ondemand_powersave_bias_init_cpu(i);
 218         }
 219 }
 220
 221 /************************** sysfs interface ************************/
 222
 223 static ssize_t show_sampling_rate_max(struct kobject *kobj,
 224                                       struct attribute *attr, char *buf)
 225 {
 226         printk_once(KERN_INFO "CPUFREQ: ondemand sampling_rate_max "
 227                "sysfs file is deprecated - used by: %s\n", current->comm);
 228         return sprintf(buf, "%u\n", -1U);
 229 }
 230
 231 static ssize_t show_sampling_rate_min(struct kobject *kobj,
 232                                       struct attribute *attr, char *buf)
 233 {
 234         return sprintf(buf, "%u\n", min_sampling_rate);
 235 }
 236
 237 #define define_one_ro(_name)            \
 238 static struct global_attr _name =       \
 239 __ATTR(_name, 0444, show_##_name, NULL)
 240
 241 define_one_ro(sampling_rate_max);
 242 define_one_ro(sampling_rate_min);
 243
 244 /* cpufreq_ondemand Governor Tunables */
 245 #define show_one(file_name, object)                                     \
 246 static ssize_t show_##file_name                                         \
 247 (struct kobject *kobj, struct attribute *attr, char *buf)              \
 248 {                                                                       \
 249         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
 250 }
 251 show_one(sampling_rate, sampling_rate);
 252 show_one(up_threshold, up_threshold);
 253 show_one(ignore_nice_load, ignore_nice);
 254 show_one(powersave_bias, powersave_bias);
 255
 256 /*** delete after deprecation time ***/
 257
 258 #define DEPRECATION_MSG(file_name)                                      \
 259         printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "       \
 260                     "interface is deprecated - " #file_name "\n");
 261
 262 #define show_one_old(file_name)                                         \
 263 static ssize_t show_##file_name##_old                                   \
 264 (struct cpufreq_policy *unused, char *buf)                              \
 265 {                                                                       \
 266         printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "       \
 267                     "interface is deprecated - " #file_name "\n");      \
 268         return show_##file_name(NULL, NULL, buf);                       \
 269 }
 270 show_one_old(sampling_rate);
 271 show_one_old(up_threshold);
 272 show_one_old(ignore_nice_load);
 273 show_one_old(powersave_bias);
 274 show_one_old(sampling_rate_min);
 275 show_one_old(sampling_rate_max);
 276
 277 #define define_one_ro_old(object, _name)       \
 278 static struct freq_attr object =               \
 279 __ATTR(_name, 0444, show_##_name##_old, NULL)
 280
 281 define_one_ro_old(sampling_rate_min_old, sampling_rate_min);
 282 define_one_ro_old(sampling_rate_max_old, sampling_rate_max);
 283
 284 /*** delete after deprecation time ***/
 285
 286 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
 287                                    const char *buf, size_t count)
 288 {
 289         unsigned int input;
 290         int ret;
 291         ret = sscanf(buf, "%u", &input);
 292         if (ret != 1)
 293                 return -EINVAL;
 294
 295         mutex_lock(&dbs_mutex);
 296         dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
 297         mutex_unlock(&dbs_mutex);
 298
 299         return count;
 300 }
 301
 302 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
 303                                   const char *buf, size_t count)
 304 {
 305         unsigned int input;
 306         int ret;
 307         ret = sscanf(buf, "%u", &input);
 308
 309         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 310                         input < MIN_FREQUENCY_UP_THRESHOLD) {
 311                 return -EINVAL;
 312         }
 313
 314         mutex_lock(&dbs_mutex);
 315         dbs_tuners_ins.up_threshold = input;
 316         mutex_unlock(&dbs_mutex);
 317
 318         return count;
 319 }
 320
 321 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
 322                                       const char *buf, size_t count)
 323 {
 324         unsigned int input;
 325         int ret;
 326
 327         unsigned int j;
 328
 329         ret = sscanf(buf, "%u", &input);
 330         if (ret != 1)
 331                 return -EINVAL;
 332
 333         if (input > 1)
 334                 input = 1;
 335
 336         mutex_lock(&dbs_mutex);
 337         if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
 338                 mutex_unlock(&dbs_mutex);
 339                 return count;
 340         }
 341         dbs_tuners_ins.ignore_nice = input;
 342
 343         /* we need to re-evaluate prev_cpu_idle */
 344         for_each_online_cpu(j) {
 345                 struct cpu_dbs_info_s *dbs_info;
 346                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
 347                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 348                                                 &dbs_info->prev_cpu_wall);
 349                 if (dbs_tuners_ins.ignore_nice)
 350                         dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 351
 352         }
 353         mutex_unlock(&dbs_mutex);
 354
 355         return count;
 356 }
 357
 358 static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b,
 359                                     const char *buf, size_t count)
 360 {
 361         unsigned int input;
 362         int ret;
 363         ret = sscanf(buf, "%u", &input);
 364
 365         if (ret != 1)
 366                 return -EINVAL;
 367
 368         if (input > 1000)
 369                 input = 1000;
 370
 371         mutex_lock(&dbs_mutex);
 372         dbs_tuners_ins.powersave_bias = input;
 373         ondemand_powersave_bias_init();
 374         mutex_unlock(&dbs_mutex);
 375
 376         return count;
 377 }
 378
 379 #define define_one_rw(_name) \
 380 static struct global_attr _name = \
 381 __ATTR(_name, 0644, show_##_name, store_##_name)
 382
 383 define_one_rw(sampling_rate);
 384 define_one_rw(up_threshold);
 385 define_one_rw(ignore_nice_load);
 386 define_one_rw(powersave_bias);
 387
 388 static struct attribute *dbs_attributes[] = {
 389         &sampling_rate_max.attr,
 390         &sampling_rate_min.attr,
 391         &sampling_rate.attr,
 392         &up_threshold.attr,
 393         &ignore_nice_load.attr,
 394         &powersave_bias.attr,
 395         NULL
 396 };
 397
 398 static struct attribute_group dbs_attr_group = {
 399         .attrs = dbs_attributes,
 400         .name = "ondemand",
 401 };
 402
 403 /*** delete after deprecation time ***/
 404
 405 #define write_one_old(file_name)                                        \
 406 static ssize_t store_##file_name##_old                                  \
 407 (struct cpufreq_policy *unused, const char *buf, size_t count)          \
 408 {                                                                       \
 409        printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "        \
 410                    "interface is deprecated - " #file_name "\n");       \
 411        return store_##file_name(NULL, NULL, buf, count);                \
 412 }
 413 write_one_old(sampling_rate);
 414 write_one_old(up_threshold);
 415 write_one_old(ignore_nice_load);
 416 write_one_old(powersave_bias);
 417
 418 #define define_one_rw_old(object, _name)       \
 419 static struct freq_attr object =               \
 420 __ATTR(_name, 0644, show_##_name##_old, store_##_name##_old)
 421
 422 define_one_rw_old(sampling_rate_old, sampling_rate);
 423 define_one_rw_old(up_threshold_old, up_threshold);
 424 define_one_rw_old(ignore_nice_load_old, ignore_nice_load);
 425 define_one_rw_old(powersave_bias_old, powersave_bias);
 426
 427 static struct attribute *dbs_attributes_old[] = {
 428        &sampling_rate_max_old.attr,
 429        &sampling_rate_min_old.attr,
 430        &sampling_rate_old.attr,
 431        &up_threshold_old.attr,
 432        &ignore_nice_load_old.attr,
 433        &powersave_bias_old.attr,
 434        NULL
 435 };
 436
 437 static struct attribute_group dbs_attr_group_old = {
 438        .attrs = dbs_attributes_old,
 439        .name = "ondemand",
 440 };
 441
 442 /*** delete after deprecation time ***/
 443
 444 /************************** sysfs end ************************/
 445
 446 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
 447 {
 448         unsigned int max_load_freq;
 449
 450         struct cpufreq_policy *policy;
 451         unsigned int j;
 452
 453         this_dbs_info->freq_lo = 0;
 454         policy = this_dbs_info->cur_policy;
 455
 456         /*
 457          * Every sampling_rate, we check, if current idle time is less
 458          * than 20% (default), then we try to increase frequency
 459          * Every sampling_rate, we look for a the lowest
 460          * frequency which can sustain the load while keeping idle time over
 461          * 30%. If such a frequency exist, we try to decrease to this frequency.
 462          *
 463          * Any frequency increase takes it to the maximum frequency.
 464          * Frequency reduction happens at minimum steps of
 465          * 5% (default) of current frequency
 466          */
 467
 468         /* Get Absolute Load - in terms of freq */
 469         max_load_freq = 0;
 470
 471         for_each_cpu(j, policy->cpus) {
 472                 struct cpu_dbs_info_s *j_dbs_info;
 473                 cputime64_t cur_wall_time, cur_idle_time;
 474                 unsigned int idle_time, wall_time;
 475                 unsigned int load, load_freq;
 476                 int freq_avg;
 477
 478                 j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
 479
 480                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
 481
 482                 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
 483                                 j_dbs_info->prev_cpu_wall);
 484                 j_dbs_info->prev_cpu_wall = cur_wall_time;
 485
 486                 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
 487                                 j_dbs_info->prev_cpu_idle);
 488                 j_dbs_info->prev_cpu_idle = cur_idle_time;
 489
 490                 if (dbs_tuners_ins.ignore_nice) {
 491                         cputime64_t cur_nice;
 492                         unsigned long cur_nice_jiffies;
 493
 494                         cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
 495                                          j_dbs_info->prev_cpu_nice);
 496                         /*
 497                          * Assumption: nice time between sampling periods will
 498                          * be less than 2^32 jiffies for 32 bit sys
 499                          */
 500                         cur_nice_jiffies = (unsigned long)
 501                                         cputime64_to_jiffies64(cur_nice);
 502
 503                         j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 504                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
 505                 }
 506
 507                 if (unlikely(!wall_time || wall_time < idle_time))
 508                         continue;
 509
 510                 load = 100 * (wall_time - idle_time) / wall_time;
 511
 512                 freq_avg = __cpufreq_driver_getavg(policy, j);
 513                 if (freq_avg <= 0)
 514                         freq_avg = policy->cur;
 515
 516                 load_freq = load * freq_avg;
 517                 if (load_freq > max_load_freq)
 518                         max_load_freq = load_freq;
 519         }
 520
 521         /* Check for frequency increase */
 522         if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
 523                 /* if we are already at full speed then break out early */
 524                 if (!dbs_tuners_ins.powersave_bias) {
 525                         if (policy->cur == policy->max)
 526                                 return;
 527
 528                         __cpufreq_driver_target(policy, policy->max,
 529                                 CPUFREQ_RELATION_H);
 530                 } else {
 531                         int freq = powersave_bias_target(policy, policy->max,
 532                                         CPUFREQ_RELATION_H);
 533                         __cpufreq_driver_target(policy, freq,
 534                                 CPUFREQ_RELATION_L);
 535                 }
 536                 return;
 537         }
 538
 539         /* Check for frequency decrease */
 540         /* if we cannot reduce the frequency anymore, break out early */
 541         if (policy->cur == policy->min)
 542                 return;
 543
 544         /*
 545          * The optimal frequency is the frequency that is the lowest that
 546          * can support the current CPU usage without triggering the up
 547          * policy. To be safe, we focus 10 points under the threshold.
 548          */
 549         if (max_load_freq <
 550             (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
 551              policy->cur) {
 552                 unsigned int freq_next;
 553                 freq_next = max_load_freq /
 554                                 (dbs_tuners_ins.up_threshold -
 555                                  dbs_tuners_ins.down_differential);
 556
 557                 if (!dbs_tuners_ins.powersave_bias) {
 558                         __cpufreq_driver_target(policy, freq_next,
 559                                         CPUFREQ_RELATION_L);
 560                 } else {
 561                         int freq = powersave_bias_target(policy, freq_next,
 562                                         CPUFREQ_RELATION_L);
 563                         __cpufreq_driver_target(policy, freq,
 564                                 CPUFREQ_RELATION_L);
 565                 }
 566         }
 567 }
 568
 569 static void do_dbs_timer(struct work_struct *work)
 570 {
 571         struct cpu_dbs_info_s *dbs_info =
 572                 container_of(work, struct cpu_dbs_info_s, work.work);
 573         unsigned int cpu = dbs_info->cpu;
 574         int sample_type = dbs_info->sample_type;
 575
 576         /* We want all CPUs to do sampling nearly on same jiffy */
 577         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 578
 579         if (num_online_cpus() > 1)
 580                 delay -= jiffies % delay;
 581
 582         mutex_lock(&dbs_info->timer_mutex);
 583
 584         /* Common NORMAL_SAMPLE setup */
 585         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 586         if (!dbs_tuners_ins.powersave_bias ||
 587             sample_type == DBS_NORMAL_SAMPLE) {
 588                 dbs_check_cpu(dbs_info);
 589                 if (dbs_info->freq_lo) {
 590                         /* Setup timer for SUB_SAMPLE */
 591                         dbs_info->sample_type = DBS_SUB_SAMPLE;
 592                         delay = dbs_info->freq_hi_jiffies;
 593                 }
 594         } else {
 595                 __cpufreq_driver_target(dbs_info->cur_policy,
 596                         dbs_info->freq_lo, CPUFREQ_RELATION_H);
 597         }
 598         queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
 599         mutex_unlock(&dbs_info->timer_mutex);
 600 }
 601
 602 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
 603 {
 604         /* We want all CPUs to do sampling nearly on same jiffy */
 605         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 606         delay -= jiffies % delay;
 607
 608         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 609         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
 610         queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work,
 611                 delay);
 612 }
 613
 614 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
 615 {
 616         cancel_delayed_work_sync(&dbs_info->work);
 617 }
 618
 619 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 620                                    unsigned int event)
 621 {
 622         unsigned int cpu = policy->cpu;
 623         struct cpu_dbs_info_s *this_dbs_info;
 624         unsigned int j;
 625         int rc;
 626
 627         this_dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 628
 629         switch (event) {
 630         case CPUFREQ_GOV_START:
 631                 if ((!cpu_online(cpu)) || (!policy->cur))
 632                         return -EINVAL;
 633
 634                 mutex_lock(&dbs_mutex);
 635
 636                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group_old);
 637                 if (rc) {
 638                         mutex_unlock(&dbs_mutex);
 639                         return rc;
 640                 }
 641
 642                 dbs_enable++;
 643                 for_each_cpu(j, policy->cpus) {
 644                         struct cpu_dbs_info_s *j_dbs_info;
 645                         j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
 646                         j_dbs_info->cur_policy = policy;
 647
 648                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 649                                                 &j_dbs_info->prev_cpu_wall);
 650                         if (dbs_tuners_ins.ignore_nice) {
 651                                 j_dbs_info->prev_cpu_nice =
 652                                                 kstat_cpu(j).cpustat.nice;
 653                         }
 654                 }
 655                 this_dbs_info->cpu = cpu;
 656                 ondemand_powersave_bias_init_cpu(cpu);
 657                 /*
 658                  * Start the timerschedule work, when this governor
 659                  * is used for first time
 660                  */
 661                 if (dbs_enable == 1) {
 662                         unsigned int latency;
 663
 664                         rc = sysfs_create_group(cpufreq_global_kobject,
 665                                                 &dbs_attr_group);
 666                         if (rc) {
 667                                 mutex_unlock(&dbs_mutex);
 668                                 return rc;
 669                         }
 670
 671                         /* policy latency is in nS. Convert it to uS first */
 672                         latency = policy->cpuinfo.transition_latency / 1000;
 673                         if (latency == 0)
 674                                 latency = 1;
 675                         /* Bring kernel and HW constraints together */
 676                         min_sampling_rate = max(min_sampling_rate,
 677                                         MIN_LATENCY_MULTIPLIER * latency);
 678                         dbs_tuners_ins.sampling_rate =
 679                                 max(min_sampling_rate,
 680                                     latency * LATENCY_MULTIPLIER);
 681                 }
 682                 mutex_unlock(&dbs_mutex);
 683
 684                 mutex_init(&this_dbs_info->timer_mutex);
 685                 dbs_timer_init(this_dbs_info);
 686                 break;
 687
 688         case CPUFREQ_GOV_STOP:
 689                 dbs_timer_exit(this_dbs_info);
 690
 691                 mutex_lock(&dbs_mutex);
 692                 sysfs_remove_group(&policy->kobj, &dbs_attr_group_old);
 693                 mutex_destroy(&this_dbs_info->timer_mutex);
 694                 dbs_enable--;
 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(this_dbs_info->cur_policy,
 706                                 policy->max, CPUFREQ_RELATION_H);
 707                 else if (policy->min > this_dbs_info->cur_policy->cur)
 708                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 709                                 policy->min, CPUFREQ_RELATION_L);
 710                 mutex_unlock(&this_dbs_info->timer_mutex);
 711                 break;
 712         }
 713         return 0;
 714 }
 715
 716 static int __init cpufreq_gov_dbs_init(void)
 717 {
 718         int err;
 719         cputime64_t wall;
 720         u64 idle_time;
 721         int cpu = get_cpu();
 722
 723         idle_time = get_cpu_idle_time_us(cpu, &wall);
 724         put_cpu();
 725         if (idle_time != -1ULL) {
 726                 /* Idle micro accounting is supported. Use finer thresholds */
 727                 dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
 728                 dbs_tuners_ins.down_differential =
 729                                         MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
 730                 /*
 731                  * In no_hz/micro accounting case we set the minimum frequency
 732                  * not depending on HZ, but fixed (very low). The deferred
 733                  * timer might skip some samples if idle/sleeping as needed.
 734                 */
 735                 min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
 736         } else {
 737                 /* For correct statistics, we need 10 ticks for each measure */
 738                 min_sampling_rate =
 739                         MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
 740         }
 741
 742         kondemand_wq = create_workqueue("kondemand");
 743         if (!kondemand_wq) {
 744                 printk(KERN_ERR "Creation of kondemand failed\n");
 745                 return -EFAULT;
 746         }
 747         err = cpufreq_register_governor(&cpufreq_gov_ondemand);
 748         if (err)
 749                 destroy_workqueue(kondemand_wq);
 750
 751         return err;
 752 }
 753
 754 static void __exit cpufreq_gov_dbs_exit(void)
 755 {
 756         cpufreq_unregister_governor(&cpufreq_gov_ondemand);
 757         destroy_workqueue(kondemand_wq);
 758 }
 759
 760
 761 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 762 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 763 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 764         "Low Latency Frequency Transition capable processors");
 765 MODULE_LICENSE("GPL");
 766
 767 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 768 fs_initcall(cpufreq_gov_dbs_init);
 769 #else
 770 module_init(cpufreq_gov_dbs_init);
 771 #endif
 772 module_exit(cpufreq_gov_dbs_exit);