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_one_global_ro(sampling_rate_max);
 238 define_one_global_ro(sampling_rate_min);
 239
 240 /* cpufreq_ondemand Governor Tunables */
 241 #define show_one(file_name, object)                                     \
 242 static ssize_t show_##file_name                                         \
 243 (struct kobject *kobj, struct attribute *attr, char *buf)              \
 244 {                                                                       \
 245         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
 246 }
 247 show_one(sampling_rate, sampling_rate);
 248 show_one(up_threshold, up_threshold);
 249 show_one(ignore_nice_load, ignore_nice);
 250 show_one(powersave_bias, powersave_bias);
 251
 252 /*** delete after deprecation time ***/
 253
 254 #define DEPRECATION_MSG(file_name)                                      \
 255         printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "       \
 256                     "interface is deprecated - " #file_name "\n");
 257
 258 #define show_one_old(file_name)                                         \
 259 static ssize_t show_##file_name##_old                                   \
 260 (struct cpufreq_policy *unused, char *buf)                              \
 261 {                                                                       \
 262         printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "       \
 263                     "interface is deprecated - " #file_name "\n");      \
 264         return show_##file_name(NULL, NULL, buf);                       \
 265 }
 266 show_one_old(sampling_rate);
 267 show_one_old(up_threshold);
 268 show_one_old(ignore_nice_load);
 269 show_one_old(powersave_bias);
 270 show_one_old(sampling_rate_min);
 271 show_one_old(sampling_rate_max);
 272
 273 cpufreq_freq_attr_ro_old(sampling_rate_min);
 274 cpufreq_freq_attr_ro_old(sampling_rate_max);
 275
 276 /*** delete after deprecation time ***/
 277
 278 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
 279                                    const char *buf, size_t count)
 280 {
 281         unsigned int input;
 282         int ret;
 283         ret = sscanf(buf, "%u", &input);
 284         if (ret != 1)
 285                 return -EINVAL;
 286
 287         mutex_lock(&dbs_mutex);
 288         dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
 289         mutex_unlock(&dbs_mutex);
 290
 291         return count;
 292 }
 293
 294 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
 295                                   const char *buf, size_t count)
 296 {
 297         unsigned int input;
 298         int ret;
 299         ret = sscanf(buf, "%u", &input);
 300
 301         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 302                         input < MIN_FREQUENCY_UP_THRESHOLD) {
 303                 return -EINVAL;
 304         }
 305
 306         mutex_lock(&dbs_mutex);
 307         dbs_tuners_ins.up_threshold = input;
 308         mutex_unlock(&dbs_mutex);
 309
 310         return count;
 311 }
 312
 313 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
 314                                       const char *buf, size_t count)
 315 {
 316         unsigned int input;
 317         int ret;
 318
 319         unsigned int j;
 320
 321         ret = sscanf(buf, "%u", &input);
 322         if (ret != 1)
 323                 return -EINVAL;
 324
 325         if (input > 1)
 326                 input = 1;
 327
 328         mutex_lock(&dbs_mutex);
 329         if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
 330                 mutex_unlock(&dbs_mutex);
 331                 return count;
 332         }
 333         dbs_tuners_ins.ignore_nice = input;
 334
 335         /* we need to re-evaluate prev_cpu_idle */
 336         for_each_online_cpu(j) {
 337                 struct cpu_dbs_info_s *dbs_info;
 338                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
 339                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 340                                                 &dbs_info->prev_cpu_wall);
 341                 if (dbs_tuners_ins.ignore_nice)
 342                         dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 343
 344         }
 345         mutex_unlock(&dbs_mutex);
 346
 347         return count;
 348 }
 349
 350 static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b,
 351                                     const char *buf, size_t count)
 352 {
 353         unsigned int input;
 354         int ret;
 355         ret = sscanf(buf, "%u", &input);
 356
 357         if (ret != 1)
 358                 return -EINVAL;
 359
 360         if (input > 1000)
 361                 input = 1000;
 362
 363         mutex_lock(&dbs_mutex);
 364         dbs_tuners_ins.powersave_bias = input;
 365         ondemand_powersave_bias_init();
 366         mutex_unlock(&dbs_mutex);
 367
 368         return count;
 369 }
 370
 371 define_one_global_rw(sampling_rate);
 372 define_one_global_rw(up_threshold);
 373 define_one_global_rw(ignore_nice_load);
 374 define_one_global_rw(powersave_bias);
 375
 376 static struct attribute *dbs_attributes[] = {
 377         &sampling_rate_max.attr,
 378         &sampling_rate_min.attr,
 379         &sampling_rate.attr,
 380         &up_threshold.attr,
 381         &ignore_nice_load.attr,
 382         &powersave_bias.attr,
 383         NULL
 384 };
 385
 386 static struct attribute_group dbs_attr_group = {
 387         .attrs = dbs_attributes,
 388         .name = "ondemand",
 389 };
 390
 391 /*** delete after deprecation time ***/
 392
 393 #define write_one_old(file_name)                                        \
 394 static ssize_t store_##file_name##_old                                  \
 395 (struct cpufreq_policy *unused, const char *buf, size_t count)          \
 396 {                                                                       \
 397        printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "        \
 398                    "interface is deprecated - " #file_name "\n");       \
 399        return store_##file_name(NULL, NULL, buf, count);                \
 400 }
 401 write_one_old(sampling_rate);
 402 write_one_old(up_threshold);
 403 write_one_old(ignore_nice_load);
 404 write_one_old(powersave_bias);
 405
 406 cpufreq_freq_attr_rw_old(sampling_rate);
 407 cpufreq_freq_attr_rw_old(up_threshold);
 408 cpufreq_freq_attr_rw_old(ignore_nice_load);
 409 cpufreq_freq_attr_rw_old(powersave_bias);
 410
 411 static struct attribute *dbs_attributes_old[] = {
 412        &sampling_rate_max_old.attr,
 413        &sampling_rate_min_old.attr,
 414        &sampling_rate_old.attr,
 415        &up_threshold_old.attr,
 416        &ignore_nice_load_old.attr,
 417        &powersave_bias_old.attr,
 418        NULL
 419 };
 420
 421 static struct attribute_group dbs_attr_group_old = {
 422        .attrs = dbs_attributes_old,
 423        .name = "ondemand",
 424 };
 425
 426 /*** delete after deprecation time ***/
 427
 428 /************************** sysfs end ************************/
 429
 430 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
 431 {
 432         unsigned int max_load_freq;
 433
 434         struct cpufreq_policy *policy;
 435         unsigned int j;
 436
 437         this_dbs_info->freq_lo = 0;
 438         policy = this_dbs_info->cur_policy;
 439
 440         /*
 441          * Every sampling_rate, we check, if current idle time is less
 442          * than 20% (default), then we try to increase frequency
 443          * Every sampling_rate, we look for a the lowest
 444          * frequency which can sustain the load while keeping idle time over
 445          * 30%. If such a frequency exist, we try to decrease to this frequency.
 446          *
 447          * Any frequency increase takes it to the maximum frequency.
 448          * Frequency reduction happens at minimum steps of
 449          * 5% (default) of current frequency
 450          */
 451
 452         /* Get Absolute Load - in terms of freq */
 453         max_load_freq = 0;
 454
 455         for_each_cpu(j, policy->cpus) {
 456                 struct cpu_dbs_info_s *j_dbs_info;
 457                 cputime64_t cur_wall_time, cur_idle_time;
 458                 unsigned int idle_time, wall_time;
 459                 unsigned int load, load_freq;
 460                 int freq_avg;
 461
 462                 j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
 463
 464                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
 465
 466                 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
 467                                 j_dbs_info->prev_cpu_wall);
 468                 j_dbs_info->prev_cpu_wall = cur_wall_time;
 469
 470                 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
 471                                 j_dbs_info->prev_cpu_idle);
 472                 j_dbs_info->prev_cpu_idle = cur_idle_time;
 473
 474                 if (dbs_tuners_ins.ignore_nice) {
 475                         cputime64_t cur_nice;
 476                         unsigned long cur_nice_jiffies;
 477
 478                         cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
 479                                          j_dbs_info->prev_cpu_nice);
 480                         /*
 481                          * Assumption: nice time between sampling periods will
 482                          * be less than 2^32 jiffies for 32 bit sys
 483                          */
 484                         cur_nice_jiffies = (unsigned long)
 485                                         cputime64_to_jiffies64(cur_nice);
 486
 487                         j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 488                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
 489                 }
 490
 491                 if (unlikely(!wall_time || wall_time < idle_time))
 492                         continue;
 493
 494                 load = 100 * (wall_time - idle_time) / wall_time;
 495
 496                 freq_avg = __cpufreq_driver_getavg(policy, j);
 497                 if (freq_avg <= 0)
 498                         freq_avg = policy->cur;
 499
 500                 load_freq = load * freq_avg;
 501                 if (load_freq > max_load_freq)
 502                         max_load_freq = load_freq;
 503         }
 504
 505         /* Check for frequency increase */
 506         if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
 507                 /* if we are already at full speed then break out early */
 508                 if (!dbs_tuners_ins.powersave_bias) {
 509                         if (policy->cur == policy->max)
 510                                 return;
 511
 512                         __cpufreq_driver_target(policy, policy->max,
 513                                 CPUFREQ_RELATION_H);
 514                 } else {
 515                         int freq = powersave_bias_target(policy, policy->max,
 516                                         CPUFREQ_RELATION_H);
 517                         __cpufreq_driver_target(policy, freq,
 518                                 CPUFREQ_RELATION_L);
 519                 }
 520                 return;
 521         }
 522
 523         /* Check for frequency decrease */
 524         /* if we cannot reduce the frequency anymore, break out early */
 525         if (policy->cur == policy->min)
 526                 return;
 527
 528         /*
 529          * The optimal frequency is the frequency that is the lowest that
 530          * can support the current CPU usage without triggering the up
 531          * policy. To be safe, we focus 10 points under the threshold.
 532          */
 533         if (max_load_freq <
 534             (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
 535              policy->cur) {
 536                 unsigned int freq_next;
 537                 freq_next = max_load_freq /
 538                                 (dbs_tuners_ins.up_threshold -
 539                                  dbs_tuners_ins.down_differential);
 540
 541                 if (freq_next < policy->min)
 542                         freq_next = policy->min;
 543
 544                 if (!dbs_tuners_ins.powersave_bias) {
 545                         __cpufreq_driver_target(policy, freq_next,
 546                                         CPUFREQ_RELATION_L);
 547                 } else {
 548                         int freq = powersave_bias_target(policy, freq_next,
 549                                         CPUFREQ_RELATION_L);
 550                         __cpufreq_driver_target(policy, freq,
 551                                 CPUFREQ_RELATION_L);
 552                 }
 553         }
 554 }
 555
 556 static void do_dbs_timer(struct work_struct *work)
 557 {
 558         struct cpu_dbs_info_s *dbs_info =
 559                 container_of(work, struct cpu_dbs_info_s, work.work);
 560         unsigned int cpu = dbs_info->cpu;
 561         int sample_type = dbs_info->sample_type;
 562
 563         /* We want all CPUs to do sampling nearly on same jiffy */
 564         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 565
 566         delay -= jiffies % delay;
 567         mutex_lock(&dbs_info->timer_mutex);
 568
 569         /* Common NORMAL_SAMPLE setup */
 570         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 571         if (!dbs_tuners_ins.powersave_bias ||
 572             sample_type == DBS_NORMAL_SAMPLE) {
 573                 dbs_check_cpu(dbs_info);
 574                 if (dbs_info->freq_lo) {
 575                         /* Setup timer for SUB_SAMPLE */
 576                         dbs_info->sample_type = DBS_SUB_SAMPLE;
 577                         delay = dbs_info->freq_hi_jiffies;
 578                 }
 579         } else {
 580                 __cpufreq_driver_target(dbs_info->cur_policy,
 581                         dbs_info->freq_lo, CPUFREQ_RELATION_H);
 582         }
 583         queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
 584         mutex_unlock(&dbs_info->timer_mutex);
 585 }
 586
 587 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
 588 {
 589         /* We want all CPUs to do sampling nearly on same jiffy */
 590         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 591         delay -= jiffies % delay;
 592
 593         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 594         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
 595         queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work,
 596                 delay);
 597 }
 598
 599 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
 600 {
 601         cancel_delayed_work_sync(&dbs_info->work);
 602 }
 603
 604 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 605                                    unsigned int event)
 606 {
 607         unsigned int cpu = policy->cpu;
 608         struct cpu_dbs_info_s *this_dbs_info;
 609         unsigned int j;
 610         int rc;
 611
 612         this_dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 613
 614         switch (event) {
 615         case CPUFREQ_GOV_START:
 616                 if ((!cpu_online(cpu)) || (!policy->cur))
 617                         return -EINVAL;
 618
 619                 mutex_lock(&dbs_mutex);
 620
 621                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group_old);
 622                 if (rc) {
 623                         mutex_unlock(&dbs_mutex);
 624                         return rc;
 625                 }
 626
 627                 dbs_enable++;
 628                 for_each_cpu(j, policy->cpus) {
 629                         struct cpu_dbs_info_s *j_dbs_info;
 630                         j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
 631                         j_dbs_info->cur_policy = policy;
 632
 633                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 634                                                 &j_dbs_info->prev_cpu_wall);
 635                         if (dbs_tuners_ins.ignore_nice) {
 636                                 j_dbs_info->prev_cpu_nice =
 637                                                 kstat_cpu(j).cpustat.nice;
 638                         }
 639                 }
 640                 this_dbs_info->cpu = cpu;
 641                 ondemand_powersave_bias_init_cpu(cpu);
 642                 /*
 643                  * Start the timerschedule work, when this governor
 644                  * is used for first time
 645                  */
 646                 if (dbs_enable == 1) {
 647                         unsigned int latency;
 648
 649                         rc = sysfs_create_group(cpufreq_global_kobject,
 650                                                 &dbs_attr_group);
 651                         if (rc) {
 652                                 mutex_unlock(&dbs_mutex);
 653                                 return rc;
 654                         }
 655
 656                         /* policy latency is in nS. Convert it to uS first */
 657                         latency = policy->cpuinfo.transition_latency / 1000;
 658                         if (latency == 0)
 659                                 latency = 1;
 660                         /* Bring kernel and HW constraints together */
 661                         min_sampling_rate = max(min_sampling_rate,
 662                                         MIN_LATENCY_MULTIPLIER * latency);
 663                         dbs_tuners_ins.sampling_rate =
 664                                 max(min_sampling_rate,
 665                                     latency * LATENCY_MULTIPLIER);
 666                 }
 667                 mutex_unlock(&dbs_mutex);
 668
 669                 mutex_init(&this_dbs_info->timer_mutex);
 670                 dbs_timer_init(this_dbs_info);
 671                 break;
 672
 673         case CPUFREQ_GOV_STOP:
 674                 dbs_timer_exit(this_dbs_info);
 675
 676                 mutex_lock(&dbs_mutex);
 677                 sysfs_remove_group(&policy->kobj, &dbs_attr_group_old);
 678                 mutex_destroy(&this_dbs_info->timer_mutex);
 679                 dbs_enable--;
 680                 mutex_unlock(&dbs_mutex);
 681                 if (!dbs_enable)
 682                         sysfs_remove_group(cpufreq_global_kobject,
 683                                            &dbs_attr_group);
 684
 685                 break;
 686
 687         case CPUFREQ_GOV_LIMITS:
 688                 mutex_lock(&this_dbs_info->timer_mutex);
 689                 if (policy->max < this_dbs_info->cur_policy->cur)
 690                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 691                                 policy->max, CPUFREQ_RELATION_H);
 692                 else if (policy->min > this_dbs_info->cur_policy->cur)
 693                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 694                                 policy->min, CPUFREQ_RELATION_L);
 695                 mutex_unlock(&this_dbs_info->timer_mutex);
 696                 break;
 697         }
 698         return 0;
 699 }
 700
 701 static int __init cpufreq_gov_dbs_init(void)
 702 {
 703         int err;
 704         cputime64_t wall;
 705         u64 idle_time;
 706         int cpu = get_cpu();
 707
 708         idle_time = get_cpu_idle_time_us(cpu, &wall);
 709         put_cpu();
 710         if (idle_time != -1ULL) {
 711                 /* Idle micro accounting is supported. Use finer thresholds */
 712                 dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
 713                 dbs_tuners_ins.down_differential =
 714                                         MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
 715                 /*
 716                  * In no_hz/micro accounting case we set the minimum frequency
 717                  * not depending on HZ, but fixed (very low). The deferred
 718                  * timer might skip some samples if idle/sleeping as needed.
 719                 */
 720                 min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
 721         } else {
 722                 /* For correct statistics, we need 10 ticks for each measure */
 723                 min_sampling_rate =
 724                         MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
 725         }
 726
 727         kondemand_wq = create_workqueue("kondemand");
 728         if (!kondemand_wq) {
 729                 printk(KERN_ERR "Creation of kondemand failed\n");
 730                 return -EFAULT;
 731         }
 732         err = cpufreq_register_governor(&cpufreq_gov_ondemand);
 733         if (err)
 734                 destroy_workqueue(kondemand_wq);
 735
 736         return err;
 737 }
 738
 739 static void __exit cpufreq_gov_dbs_exit(void)
 740 {
 741         cpufreq_unregister_governor(&cpufreq_gov_ondemand);
 742         destroy_workqueue(kondemand_wq);
 743 }
 744
 745
 746 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 747 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 748 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 749         "Low Latency Frequency Transition capable processors");
 750 MODULE_LICENSE("GPL");
 751
 752 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 753 fs_initcall(cpufreq_gov_dbs_init);
 754 #else
 755 module_init(cpufreq_gov_dbs_init);
 756 #endif
 757 module_exit(cpufreq_gov_dbs_exit);