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 #ifdef CONFIG_ARCH_RK29
  34 #define MICRO_FREQUENCY_DOWN_DIFFERENTIAL       (10)
  35 #define MICRO_FREQUENCY_UP_THRESHOLD            (80)
  36 #else
  37 #define MICRO_FREQUENCY_DOWN_DIFFERENTIAL       (3)
  38 #define MICRO_FREQUENCY_UP_THRESHOLD            (95)
  39 #endif
  40 #define MICRO_FREQUENCY_MIN_SAMPLE_RATE         (10000)
  41 #define MIN_FREQUENCY_UP_THRESHOLD              (11)
  42 #define MAX_FREQUENCY_UP_THRESHOLD              (100)
  43
  44 /*
  45  * The polling frequency of this governor depends on the capability of
  46  * the processor. Default polling frequency is 1000 times the transition
  47  * latency of the processor. The governor will work on any processor with
  48  * transition latency <= 10mS, using appropriate sampling
  49  * rate.
  50  * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
  51  * this governor will not work.
  52  * All times here are in uS.
  53  */
  54 #define MIN_SAMPLING_RATE_RATIO                 (2)
  55
  56 static unsigned int min_sampling_rate;
  57
  58 #define LATENCY_MULTIPLIER                      (1000)
  59 #define MIN_LATENCY_MULTIPLIER                  (100)
  60 #define TRANSITION_LATENCY_LIMIT                (10 * 1000 * 1000)
  61
  62 static void do_dbs_timer(struct work_struct *work);
  63 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
  64                                 unsigned int event);
  65
  66 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
  67 static
  68 #endif
  69 struct cpufreq_governor cpufreq_gov_ondemand = {
  70        .name                   = "ondemand",
  71        .governor               = cpufreq_governor_dbs,
  72        .max_transition_latency = TRANSITION_LATENCY_LIMIT,
  73        .owner                  = THIS_MODULE,
  74 };
  75
  76 /* Sampling types */
  77 enum {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE};
  78
  79 struct cpu_dbs_info_s {
  80         cputime64_t prev_cpu_idle;
  81         cputime64_t prev_cpu_wall;
  82         cputime64_t prev_cpu_nice;
  83         struct cpufreq_policy *cur_policy;
  84         struct delayed_work work;
  85         struct cpufreq_frequency_table *freq_table;
  86         unsigned int freq_lo;
  87         unsigned int freq_lo_jiffies;
  88         unsigned int freq_hi_jiffies;
  89         int cpu;
  90         unsigned int sample_type:1;
  91         /*
  92          * percpu mutex that serializes governor limit change with
  93          * do_dbs_timer invocation. We do not want do_dbs_timer to run
  94          * when user is changing the governor or limits.
  95          */
  96         struct mutex timer_mutex;
  97 };
  98 static DEFINE_PER_CPU(struct cpu_dbs_info_s, od_cpu_dbs_info);
  99
 100 static unsigned int dbs_enable; /* number of CPUs using this policy */
 101
 102 /*
 103  * dbs_mutex protects data in dbs_tuners_ins from concurrent changes on
 104  * different CPUs. It protects dbs_enable in governor start/stop.
 105  */
 106 static DEFINE_MUTEX(dbs_mutex);
 107
 108 static struct workqueue_struct  *kondemand_wq;
 109
 110 static struct dbs_tuners {
 111         unsigned int sampling_rate;
 112         unsigned int up_threshold;
 113         unsigned int down_differential;
 114         unsigned int ignore_nice;
 115         unsigned int powersave_bias;
 116 } dbs_tuners_ins = {
 117         .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
 118         .down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL,
 119 #ifdef CONFIG_ARCH_RK29
 120         .ignore_nice = 1,
 121 #else
 122         .ignore_nice = 0,
 123 #endif
 124         .powersave_bias = 0,
 125 };
 126
 127 static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
 128                                                         cputime64_t *wall)
 129 {
 130         cputime64_t idle_time;
 131         cputime64_t cur_wall_time;
 132         cputime64_t busy_time;
 133
 134         cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
 135         busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
 136                         kstat_cpu(cpu).cpustat.system);
 137
 138         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
 139         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
 140         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
 141         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice);
 142
 143         idle_time = cputime64_sub(cur_wall_time, busy_time);
 144         if (wall)
 145                 *wall = (cputime64_t)jiffies_to_usecs(cur_wall_time);
 146
 147         return (cputime64_t)jiffies_to_usecs(idle_time);
 148 }
 149
 150 static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
 151 {
 152         u64 idle_time = get_cpu_idle_time_us(cpu, wall);
 153
 154         if (idle_time == -1ULL)
 155                 return get_cpu_idle_time_jiffy(cpu, wall);
 156
 157         return idle_time;
 158 }
 159
 160 /*
 161  * Find right freq to be set now with powersave_bias on.
 162  * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
 163  * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
 164  */
 165 static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
 166                                           unsigned int freq_next,
 167                                           unsigned int relation)
 168 {
 169         unsigned int freq_req, freq_reduc, freq_avg;
 170         unsigned int freq_hi, freq_lo;
 171         unsigned int index = 0;
 172         unsigned int jiffies_total, jiffies_hi, jiffies_lo;
 173         struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
 174                                                    policy->cpu);
 175
 176         if (!dbs_info->freq_table) {
 177                 dbs_info->freq_lo = 0;
 178                 dbs_info->freq_lo_jiffies = 0;
 179                 return freq_next;
 180         }
 181
 182         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
 183                         relation, &index);
 184         freq_req = dbs_info->freq_table[index].frequency;
 185         freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
 186         freq_avg = freq_req - freq_reduc;
 187
 188         /* Find freq bounds for freq_avg in freq_table */
 189         index = 0;
 190         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 191                         CPUFREQ_RELATION_H, &index);
 192         freq_lo = dbs_info->freq_table[index].frequency;
 193         index = 0;
 194         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 195                         CPUFREQ_RELATION_L, &index);
 196         freq_hi = dbs_info->freq_table[index].frequency;
 197
 198         /* Find out how long we have to be in hi and lo freqs */
 199         if (freq_hi == freq_lo) {
 200                 dbs_info->freq_lo = 0;
 201                 dbs_info->freq_lo_jiffies = 0;
 202                 return freq_lo;
 203         }
 204         jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 205         jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
 206         jiffies_hi += ((freq_hi - freq_lo) / 2);
 207         jiffies_hi /= (freq_hi - freq_lo);
 208         jiffies_lo = jiffies_total - jiffies_hi;
 209         dbs_info->freq_lo = freq_lo;
 210         dbs_info->freq_lo_jiffies = jiffies_lo;
 211         dbs_info->freq_hi_jiffies = jiffies_hi;
 212         return freq_hi;
 213 }
 214
 215 static void ondemand_powersave_bias_init_cpu(int cpu)
 216 {
 217         struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 218         dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
 219         dbs_info->freq_lo = 0;
 220 }
 221
 222 static void ondemand_powersave_bias_init(void)
 223 {
 224         int i;
 225         for_each_online_cpu(i) {
 226                 ondemand_powersave_bias_init_cpu(i);
 227         }
 228 }
 229
 230 /************************** sysfs interface ************************/
 231
 232 static ssize_t show_sampling_rate_max(struct kobject *kobj,
 233                                       struct attribute *attr, char *buf)
 234 {
 235         printk_once(KERN_INFO "CPUFREQ: ondemand sampling_rate_max "
 236                "sysfs file is deprecated - used by: %s\n", current->comm);
 237         return sprintf(buf, "%u\n", -1U);
 238 }
 239
 240 static ssize_t show_sampling_rate_min(struct kobject *kobj,
 241                                       struct attribute *attr, char *buf)
 242 {
 243         return sprintf(buf, "%u\n", min_sampling_rate);
 244 }
 245
 246 #define define_one_ro(_name)            \
 247 static struct global_attr _name =       \
 248 __ATTR(_name, 0444, show_##_name, NULL)
 249
 250 define_one_ro(sampling_rate_max);
 251 define_one_ro(sampling_rate_min);
 252
 253 /* cpufreq_ondemand Governor Tunables */
 254 #define show_one(file_name, object)                                     \
 255 static ssize_t show_##file_name                                         \
 256 (struct kobject *kobj, struct attribute *attr, char *buf)              \
 257 {                                                                       \
 258         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
 259 }
 260 show_one(sampling_rate, sampling_rate);
 261 show_one(up_threshold, up_threshold);
 262 show_one(ignore_nice_load, ignore_nice);
 263 show_one(powersave_bias, powersave_bias);
 264
 265 /*** delete after deprecation time ***/
 266
 267 #define DEPRECATION_MSG(file_name)                                      \
 268         printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "       \
 269                     "interface is deprecated - " #file_name "\n");
 270
 271 #define show_one_old(file_name)                                         \
 272 static ssize_t show_##file_name##_old                                   \
 273 (struct cpufreq_policy *unused, char *buf)                              \
 274 {                                                                       \
 275         printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "       \
 276                     "interface is deprecated - " #file_name "\n");      \
 277         return show_##file_name(NULL, NULL, buf);                       \
 278 }
 279 show_one_old(sampling_rate);
 280 show_one_old(up_threshold);
 281 show_one_old(ignore_nice_load);
 282 show_one_old(powersave_bias);
 283 show_one_old(sampling_rate_min);
 284 show_one_old(sampling_rate_max);
 285
 286 #define define_one_ro_old(object, _name)       \
 287 static struct freq_attr object =               \
 288 __ATTR(_name, 0444, show_##_name##_old, NULL)
 289
 290 define_one_ro_old(sampling_rate_min_old, sampling_rate_min);
 291 define_one_ro_old(sampling_rate_max_old, sampling_rate_max);
 292
 293 /*** delete after deprecation time ***/
 294
 295 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
 296                                    const char *buf, size_t count)
 297 {
 298         unsigned int input;
 299         int ret;
 300         ret = sscanf(buf, "%u", &input);
 301         if (ret != 1)
 302                 return -EINVAL;
 303
 304         mutex_lock(&dbs_mutex);
 305         dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
 306         mutex_unlock(&dbs_mutex);
 307
 308         return count;
 309 }
 310
 311 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
 312                                   const char *buf, size_t count)
 313 {
 314         unsigned int input;
 315         int ret;
 316         ret = sscanf(buf, "%u", &input);
 317
 318         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 319                         input < MIN_FREQUENCY_UP_THRESHOLD) {
 320                 return -EINVAL;
 321         }
 322
 323         mutex_lock(&dbs_mutex);
 324         dbs_tuners_ins.up_threshold = input;
 325         mutex_unlock(&dbs_mutex);
 326
 327         return count;
 328 }
 329
 330 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
 331                                       const char *buf, size_t count)
 332 {
 333         unsigned int input;
 334         int ret;
 335
 336         unsigned int j;
 337
 338         ret = sscanf(buf, "%u", &input);
 339         if (ret != 1)
 340                 return -EINVAL;
 341
 342         if (input > 1)
 343                 input = 1;
 344
 345         mutex_lock(&dbs_mutex);
 346         if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
 347                 mutex_unlock(&dbs_mutex);
 348                 return count;
 349         }
 350         dbs_tuners_ins.ignore_nice = input;
 351
 352         /* we need to re-evaluate prev_cpu_idle */
 353         for_each_online_cpu(j) {
 354                 struct cpu_dbs_info_s *dbs_info;
 355                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
 356                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 357                                                 &dbs_info->prev_cpu_wall);
 358                 if (dbs_tuners_ins.ignore_nice)
 359                         dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 360
 361         }
 362         mutex_unlock(&dbs_mutex);
 363
 364         return count;
 365 }
 366
 367 static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b,
 368                                     const char *buf, size_t count)
 369 {
 370         unsigned int input;
 371         int ret;
 372         ret = sscanf(buf, "%u", &input);
 373
 374         if (ret != 1)
 375                 return -EINVAL;
 376
 377         if (input > 1000)
 378                 input = 1000;
 379
 380         mutex_lock(&dbs_mutex);
 381         dbs_tuners_ins.powersave_bias = input;
 382         ondemand_powersave_bias_init();
 383         mutex_unlock(&dbs_mutex);
 384
 385         return count;
 386 }
 387
 388 #define define_one_rw(_name) \
 389 static struct global_attr _name = \
 390 __ATTR(_name, 0644, show_##_name, store_##_name)
 391
 392 define_one_rw(sampling_rate);
 393 define_one_rw(up_threshold);
 394 define_one_rw(ignore_nice_load);
 395 define_one_rw(powersave_bias);
 396
 397 static struct attribute *dbs_attributes[] = {
 398         &sampling_rate_max.attr,
 399         &sampling_rate_min.attr,
 400         &sampling_rate.attr,
 401         &up_threshold.attr,
 402         &ignore_nice_load.attr,
 403         &powersave_bias.attr,
 404         NULL
 405 };
 406
 407 static struct attribute_group dbs_attr_group = {
 408         .attrs = dbs_attributes,
 409         .name = "ondemand",
 410 };
 411
 412 /*** delete after deprecation time ***/
 413
 414 #define write_one_old(file_name)                                        \
 415 static ssize_t store_##file_name##_old                                  \
 416 (struct cpufreq_policy *unused, const char *buf, size_t count)          \
 417 {                                                                       \
 418        printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "        \
 419                    "interface is deprecated - " #file_name "\n");       \
 420        return store_##file_name(NULL, NULL, buf, count);                \
 421 }
 422 write_one_old(sampling_rate);
 423 write_one_old(up_threshold);
 424 write_one_old(ignore_nice_load);
 425 write_one_old(powersave_bias);
 426
 427 #define define_one_rw_old(object, _name)       \
 428 static struct freq_attr object =               \
 429 __ATTR(_name, 0644, show_##_name##_old, store_##_name##_old)
 430
 431 define_one_rw_old(sampling_rate_old, sampling_rate);
 432 define_one_rw_old(up_threshold_old, up_threshold);
 433 define_one_rw_old(ignore_nice_load_old, ignore_nice_load);
 434 define_one_rw_old(powersave_bias_old, powersave_bias);
 435
 436 static struct attribute *dbs_attributes_old[] = {
 437        &sampling_rate_max_old.attr,
 438        &sampling_rate_min_old.attr,
 439        &sampling_rate_old.attr,
 440        &up_threshold_old.attr,
 441        &ignore_nice_load_old.attr,
 442        &powersave_bias_old.attr,
 443        NULL
 444 };
 445
 446 static struct attribute_group dbs_attr_group_old = {
 447        .attrs = dbs_attributes_old,
 448        .name = "ondemand",
 449 };
 450
 451 /*** delete after deprecation time ***/
 452
 453 /************************** sysfs end ************************/
 454
 455 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
 456 {
 457         unsigned int max_load_freq;
 458
 459         struct cpufreq_policy *policy;
 460         unsigned int j;
 461
 462         this_dbs_info->freq_lo = 0;
 463         policy = this_dbs_info->cur_policy;
 464
 465         /*
 466          * Every sampling_rate, we check, if current idle time is less
 467          * than 20% (default), then we try to increase frequency
 468          * Every sampling_rate, we look for a the lowest
 469          * frequency which can sustain the load while keeping idle time over
 470          * 30%. If such a frequency exist, we try to decrease to this frequency.
 471          *
 472          * Any frequency increase takes it to the maximum frequency.
 473          * Frequency reduction happens at minimum steps of
 474          * 5% (default) of current frequency
 475          */
 476
 477         /* Get Absolute Load - in terms of freq */
 478         max_load_freq = 0;
 479
 480         for_each_cpu(j, policy->cpus) {
 481                 struct cpu_dbs_info_s *j_dbs_info;
 482                 cputime64_t cur_wall_time, cur_idle_time;
 483                 unsigned int idle_time, wall_time;
 484                 unsigned int load, load_freq;
 485                 int freq_avg;
 486
 487                 j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
 488
 489                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
 490
 491                 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
 492                                 j_dbs_info->prev_cpu_wall);
 493                 j_dbs_info->prev_cpu_wall = cur_wall_time;
 494
 495                 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
 496                                 j_dbs_info->prev_cpu_idle);
 497                 j_dbs_info->prev_cpu_idle = cur_idle_time;
 498
 499                 if (dbs_tuners_ins.ignore_nice) {
 500                         cputime64_t cur_nice;
 501                         unsigned long cur_nice_jiffies;
 502
 503                         cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
 504                                          j_dbs_info->prev_cpu_nice);
 505                         /*
 506                          * Assumption: nice time between sampling periods will
 507                          * be less than 2^32 jiffies for 32 bit sys
 508                          */
 509                         cur_nice_jiffies = (unsigned long)
 510                                         cputime64_to_jiffies64(cur_nice);
 511
 512                         j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 513                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
 514                 }
 515
 516                 if (unlikely(!wall_time || wall_time < idle_time))
 517                         continue;
 518
 519                 load = 100 * (wall_time - idle_time) / wall_time;
 520
 521                 freq_avg = __cpufreq_driver_getavg(policy, j);
 522                 if (freq_avg <= 0)
 523                         freq_avg = policy->cur;
 524
 525                 load_freq = load * freq_avg;
 526                 if (load_freq > max_load_freq)
 527                         max_load_freq = load_freq;
 528         }
 529
 530         /* Check for frequency increase */
 531         if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
 532                 /* if we are already at full speed then break out early */
 533                 if (!dbs_tuners_ins.powersave_bias) {
 534                         if (policy->cur == policy->max)
 535                                 return;
 536
 537                         __cpufreq_driver_target(policy, policy->max,
 538                                 CPUFREQ_RELATION_H);
 539                 } else {
 540                         int freq = powersave_bias_target(policy, policy->max,
 541                                         CPUFREQ_RELATION_H);
 542                         __cpufreq_driver_target(policy, freq,
 543                                 CPUFREQ_RELATION_L);
 544                 }
 545                 return;
 546         }
 547
 548         /* Check for frequency decrease */
 549         /* if we cannot reduce the frequency anymore, break out early */
 550         if (policy->cur == policy->min)
 551                 return;
 552
 553         /*
 554          * The optimal frequency is the frequency that is the lowest that
 555          * can support the current CPU usage without triggering the up
 556          * policy. To be safe, we focus 10 points under the threshold.
 557          */
 558         if (max_load_freq <
 559             (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
 560              policy->cur) {
 561                 unsigned int freq_next;
 562                 freq_next = max_load_freq /
 563                                 (dbs_tuners_ins.up_threshold -
 564                                  dbs_tuners_ins.down_differential);
 565
 566                 if (!dbs_tuners_ins.powersave_bias) {
 567                         __cpufreq_driver_target(policy, freq_next,
 568                                         CPUFREQ_RELATION_L);
 569                 } else {
 570                         int freq = powersave_bias_target(policy, freq_next,
 571                                         CPUFREQ_RELATION_L);
 572                         __cpufreq_driver_target(policy, freq,
 573                                 CPUFREQ_RELATION_L);
 574                 }
 575         }
 576 }
 577
 578 static void do_dbs_timer(struct work_struct *work)
 579 {
 580         struct cpu_dbs_info_s *dbs_info =
 581                 container_of(work, struct cpu_dbs_info_s, work.work);
 582         unsigned int cpu = dbs_info->cpu;
 583         int sample_type = dbs_info->sample_type;
 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
 588         delay -= jiffies % delay;
 589         mutex_lock(&dbs_info->timer_mutex);
 590
 591         /* Common NORMAL_SAMPLE setup */
 592         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 593         if (!dbs_tuners_ins.powersave_bias ||
 594             sample_type == DBS_NORMAL_SAMPLE) {
 595                 dbs_check_cpu(dbs_info);
 596                 if (dbs_info->freq_lo) {
 597                         /* Setup timer for SUB_SAMPLE */
 598                         dbs_info->sample_type = DBS_SUB_SAMPLE;
 599                         delay = dbs_info->freq_hi_jiffies;
 600                 }
 601         } else {
 602                 __cpufreq_driver_target(dbs_info->cur_policy,
 603                         dbs_info->freq_lo, CPUFREQ_RELATION_H);
 604         }
 605         queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
 606         mutex_unlock(&dbs_info->timer_mutex);
 607 }
 608
 609 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
 610 {
 611         /* We want all CPUs to do sampling nearly on same jiffy */
 612         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 613         delay -= jiffies % delay;
 614
 615         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 616         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
 617         queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work,
 618                 delay);
 619 }
 620
 621 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
 622 {
 623         cancel_delayed_work_sync(&dbs_info->work);
 624 }
 625
 626 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 627                                    unsigned int event)
 628 {
 629         unsigned int cpu = policy->cpu;
 630         struct cpu_dbs_info_s *this_dbs_info;
 631         unsigned int j;
 632         int rc;
 633
 634         this_dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 635
 636         switch (event) {
 637         case CPUFREQ_GOV_START:
 638                 if ((!cpu_online(cpu)) || (!policy->cur))
 639                         return -EINVAL;
 640
 641                 mutex_lock(&dbs_mutex);
 642
 643                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group_old);
 644                 if (rc) {
 645                         mutex_unlock(&dbs_mutex);
 646                         return rc;
 647                 }
 648
 649                 dbs_enable++;
 650                 for_each_cpu(j, policy->cpus) {
 651                         struct cpu_dbs_info_s *j_dbs_info;
 652                         j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
 653                         j_dbs_info->cur_policy = policy;
 654
 655                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 656                                                 &j_dbs_info->prev_cpu_wall);
 657                         if (dbs_tuners_ins.ignore_nice) {
 658                                 j_dbs_info->prev_cpu_nice =
 659                                                 kstat_cpu(j).cpustat.nice;
 660                         }
 661                 }
 662                 this_dbs_info->cpu = cpu;
 663                 ondemand_powersave_bias_init_cpu(cpu);
 664                 /*
 665                  * Start the timerschedule work, when this governor
 666                  * is used for first time
 667                  */
 668                 if (dbs_enable == 1) {
 669                         unsigned int latency;
 670
 671                         rc = sysfs_create_group(cpufreq_global_kobject,
 672                                                 &dbs_attr_group);
 673                         if (rc) {
 674                                 mutex_unlock(&dbs_mutex);
 675                                 return rc;
 676                         }
 677
 678                         /* policy latency is in nS. Convert it to uS first */
 679                         latency = policy->cpuinfo.transition_latency / 1000;
 680                         if (latency == 0)
 681                                 latency = 1;
 682                         /* Bring kernel and HW constraints together */
 683                         min_sampling_rate = max(min_sampling_rate,
 684                                         MIN_LATENCY_MULTIPLIER * latency);
 685                         dbs_tuners_ins.sampling_rate =
 686                                 max(min_sampling_rate,
 687                                     latency * LATENCY_MULTIPLIER);
 688                 }
 689                 mutex_unlock(&dbs_mutex);
 690
 691                 mutex_init(&this_dbs_info->timer_mutex);
 692                 dbs_timer_init(this_dbs_info);
 693                 break;
 694
 695         case CPUFREQ_GOV_STOP:
 696                 dbs_timer_exit(this_dbs_info);
 697
 698                 mutex_lock(&dbs_mutex);
 699                 sysfs_remove_group(&policy->kobj, &dbs_attr_group_old);
 700                 mutex_destroy(&this_dbs_info->timer_mutex);
 701                 dbs_enable--;
 702                 mutex_unlock(&dbs_mutex);
 703                 if (!dbs_enable)
 704                         sysfs_remove_group(cpufreq_global_kobject,
 705                                            &dbs_attr_group);
 706
 707                 break;
 708
 709         case CPUFREQ_GOV_LIMITS:
 710                 mutex_lock(&this_dbs_info->timer_mutex);
 711                 if (policy->max < this_dbs_info->cur_policy->cur)
 712                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 713                                 policy->max, CPUFREQ_RELATION_H);
 714                 else if (policy->min > this_dbs_info->cur_policy->cur)
 715                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 716                                 policy->min, CPUFREQ_RELATION_L);
 717                 mutex_unlock(&this_dbs_info->timer_mutex);
 718                 break;
 719         }
 720         return 0;
 721 }
 722
 723 static int __init cpufreq_gov_dbs_init(void)
 724 {
 725         int err;
 726         cputime64_t wall;
 727         u64 idle_time;
 728         int cpu = get_cpu();
 729
 730         idle_time = get_cpu_idle_time_us(cpu, &wall);
 731         put_cpu();
 732         if (idle_time != -1ULL) {
 733                 /* Idle micro accounting is supported. Use finer thresholds */
 734                 dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
 735                 dbs_tuners_ins.down_differential =
 736                                         MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
 737                 /*
 738                  * In no_hz/micro accounting case we set the minimum frequency
 739                  * not depending on HZ, but fixed (very low). The deferred
 740                  * timer might skip some samples if idle/sleeping as needed.
 741                 */
 742                 min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
 743         } else {
 744                 /* For correct statistics, we need 10 ticks for each measure */
 745                 min_sampling_rate =
 746                         MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
 747         }
 748
 749         kondemand_wq = create_workqueue("kondemand");
 750         if (!kondemand_wq) {
 751                 printk(KERN_ERR "Creation of kondemand failed\n");
 752                 return -EFAULT;
 753         }
 754         err = cpufreq_register_governor(&cpufreq_gov_ondemand);
 755         if (err)
 756                 destroy_workqueue(kondemand_wq);
 757
 758         return err;
 759 }
 760
 761 static void __exit cpufreq_gov_dbs_exit(void)
 762 {
 763         cpufreq_unregister_governor(&cpufreq_gov_ondemand);
 764         destroy_workqueue(kondemand_wq);
 765 }
 766
 767
 768 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 769 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 770 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 771         "Low Latency Frequency Transition capable processors");
 772 MODULE_LICENSE("GPL");
 773
 774 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 775 fs_initcall(cpufreq_gov_dbs_init);
 776 #else
 777 module_init(cpufreq_gov_dbs_init);
 778 #endif
 779 module_exit(cpufreq_gov_dbs_exit);