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         .ignore_nice = 0,
 120         .powersave_bias = 0,
 121 };
 122
 123 static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
 124                                                         cputime64_t *wall)
 125 {
 126         cputime64_t idle_time;
 127         cputime64_t cur_wall_time;
 128         cputime64_t busy_time;
 129
 130         cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
 131         busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
 132                         kstat_cpu(cpu).cpustat.system);
 133
 134         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
 135         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
 136         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
 137         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice);
 138
 139         idle_time = cputime64_sub(cur_wall_time, busy_time);
 140         if (wall)
 141                 *wall = (cputime64_t)jiffies_to_usecs(cur_wall_time);
 142
 143         return (cputime64_t)jiffies_to_usecs(idle_time);
 144 }
 145
 146 static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
 147 {
 148         u64 idle_time = get_cpu_idle_time_us(cpu, wall);
 149
 150         if (idle_time == -1ULL)
 151                 return get_cpu_idle_time_jiffy(cpu, wall);
 152
 153         return idle_time;
 154 }
 155
 156 /*
 157  * Find right freq to be set now with powersave_bias on.
 158  * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
 159  * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
 160  */
 161 static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
 162                                           unsigned int freq_next,
 163                                           unsigned int relation)
 164 {
 165         unsigned int freq_req, freq_reduc, freq_avg;
 166         unsigned int freq_hi, freq_lo;
 167         unsigned int index = 0;
 168         unsigned int jiffies_total, jiffies_hi, jiffies_lo;
 169         struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
 170                                                    policy->cpu);
 171
 172         if (!dbs_info->freq_table) {
 173                 dbs_info->freq_lo = 0;
 174                 dbs_info->freq_lo_jiffies = 0;
 175                 return freq_next;
 176         }
 177
 178         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
 179                         relation, &index);
 180         freq_req = dbs_info->freq_table[index].frequency;
 181         freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
 182         freq_avg = freq_req - freq_reduc;
 183
 184         /* Find freq bounds for freq_avg in freq_table */
 185         index = 0;
 186         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 187                         CPUFREQ_RELATION_H, &index);
 188         freq_lo = dbs_info->freq_table[index].frequency;
 189         index = 0;
 190         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 191                         CPUFREQ_RELATION_L, &index);
 192         freq_hi = dbs_info->freq_table[index].frequency;
 193
 194         /* Find out how long we have to be in hi and lo freqs */
 195         if (freq_hi == freq_lo) {
 196                 dbs_info->freq_lo = 0;
 197                 dbs_info->freq_lo_jiffies = 0;
 198                 return freq_lo;
 199         }
 200         jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 201         jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
 202         jiffies_hi += ((freq_hi - freq_lo) / 2);
 203         jiffies_hi /= (freq_hi - freq_lo);
 204         jiffies_lo = jiffies_total - jiffies_hi;
 205         dbs_info->freq_lo = freq_lo;
 206         dbs_info->freq_lo_jiffies = jiffies_lo;
 207         dbs_info->freq_hi_jiffies = jiffies_hi;
 208         return freq_hi;
 209 }
 210
 211 static void ondemand_powersave_bias_init_cpu(int cpu)
 212 {
 213         struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 214         dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
 215         dbs_info->freq_lo = 0;
 216 }
 217
 218 static void ondemand_powersave_bias_init(void)
 219 {
 220         int i;
 221         for_each_online_cpu(i) {
 222                 ondemand_powersave_bias_init_cpu(i);
 223         }
 224 }
 225
 226 /************************** sysfs interface ************************/
 227
 228 static ssize_t show_sampling_rate_max(struct kobject *kobj,
 229                                       struct attribute *attr, char *buf)
 230 {
 231         printk_once(KERN_INFO "CPUFREQ: ondemand sampling_rate_max "
 232                "sysfs file is deprecated - used by: %s\n", current->comm);
 233         return sprintf(buf, "%u\n", -1U);
 234 }
 235
 236 static ssize_t show_sampling_rate_min(struct kobject *kobj,
 237                                       struct attribute *attr, char *buf)
 238 {
 239         return sprintf(buf, "%u\n", min_sampling_rate);
 240 }
 241
 242 #define define_one_ro(_name)            \
 243 static struct global_attr _name =       \
 244 __ATTR(_name, 0444, show_##_name, NULL)
 245
 246 define_one_ro(sampling_rate_max);
 247 define_one_ro(sampling_rate_min);
 248
 249 /* cpufreq_ondemand Governor Tunables */
 250 #define show_one(file_name, object)                                     \
 251 static ssize_t show_##file_name                                         \
 252 (struct kobject *kobj, struct attribute *attr, char *buf)              \
 253 {                                                                       \
 254         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
 255 }
 256 show_one(sampling_rate, sampling_rate);
 257 show_one(up_threshold, up_threshold);
 258 show_one(ignore_nice_load, ignore_nice);
 259 show_one(powersave_bias, powersave_bias);
 260
 261 /*** delete after deprecation time ***/
 262
 263 #define DEPRECATION_MSG(file_name)                                      \
 264         printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "       \
 265                     "interface is deprecated - " #file_name "\n");
 266
 267 #define show_one_old(file_name)                                         \
 268 static ssize_t show_##file_name##_old                                   \
 269 (struct cpufreq_policy *unused, char *buf)                              \
 270 {                                                                       \
 271         printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "       \
 272                     "interface is deprecated - " #file_name "\n");      \
 273         return show_##file_name(NULL, NULL, buf);                       \
 274 }
 275 show_one_old(sampling_rate);
 276 show_one_old(up_threshold);
 277 show_one_old(ignore_nice_load);
 278 show_one_old(powersave_bias);
 279 show_one_old(sampling_rate_min);
 280 show_one_old(sampling_rate_max);
 281
 282 #define define_one_ro_old(object, _name)       \
 283 static struct freq_attr object =               \
 284 __ATTR(_name, 0444, show_##_name##_old, NULL)
 285
 286 define_one_ro_old(sampling_rate_min_old, sampling_rate_min);
 287 define_one_ro_old(sampling_rate_max_old, sampling_rate_max);
 288
 289 /*** delete after deprecation time ***/
 290
 291 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
 292                                    const char *buf, size_t count)
 293 {
 294         unsigned int input;
 295         int ret;
 296         ret = sscanf(buf, "%u", &input);
 297         if (ret != 1)
 298                 return -EINVAL;
 299
 300         mutex_lock(&dbs_mutex);
 301         dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
 302         mutex_unlock(&dbs_mutex);
 303
 304         return count;
 305 }
 306
 307 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
 308                                   const char *buf, size_t count)
 309 {
 310         unsigned int input;
 311         int ret;
 312         ret = sscanf(buf, "%u", &input);
 313
 314         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 315                         input < MIN_FREQUENCY_UP_THRESHOLD) {
 316                 return -EINVAL;
 317         }
 318
 319         mutex_lock(&dbs_mutex);
 320         dbs_tuners_ins.up_threshold = input;
 321         mutex_unlock(&dbs_mutex);
 322
 323         return count;
 324 }
 325
 326 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
 327                                       const char *buf, size_t count)
 328 {
 329         unsigned int input;
 330         int ret;
 331
 332         unsigned int j;
 333
 334         ret = sscanf(buf, "%u", &input);
 335         if (ret != 1)
 336                 return -EINVAL;
 337
 338         if (input > 1)
 339                 input = 1;
 340
 341         mutex_lock(&dbs_mutex);
 342         if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
 343                 mutex_unlock(&dbs_mutex);
 344                 return count;
 345         }
 346         dbs_tuners_ins.ignore_nice = input;
 347
 348         /* we need to re-evaluate prev_cpu_idle */
 349         for_each_online_cpu(j) {
 350                 struct cpu_dbs_info_s *dbs_info;
 351                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
 352                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 353                                                 &dbs_info->prev_cpu_wall);
 354                 if (dbs_tuners_ins.ignore_nice)
 355                         dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 356
 357         }
 358         mutex_unlock(&dbs_mutex);
 359
 360         return count;
 361 }
 362
 363 static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b,
 364                                     const char *buf, size_t count)
 365 {
 366         unsigned int input;
 367         int ret;
 368         ret = sscanf(buf, "%u", &input);
 369
 370         if (ret != 1)
 371                 return -EINVAL;
 372
 373         if (input > 1000)
 374                 input = 1000;
 375
 376         mutex_lock(&dbs_mutex);
 377         dbs_tuners_ins.powersave_bias = input;
 378         ondemand_powersave_bias_init();
 379         mutex_unlock(&dbs_mutex);
 380
 381         return count;
 382 }
 383
 384 #define define_one_rw(_name) \
 385 static struct global_attr _name = \
 386 __ATTR(_name, 0644, show_##_name, store_##_name)
 387
 388 define_one_rw(sampling_rate);
 389 define_one_rw(up_threshold);
 390 define_one_rw(ignore_nice_load);
 391 define_one_rw(powersave_bias);
 392
 393 static struct attribute *dbs_attributes[] = {
 394         &sampling_rate_max.attr,
 395         &sampling_rate_min.attr,
 396         &sampling_rate.attr,
 397         &up_threshold.attr,
 398         &ignore_nice_load.attr,
 399         &powersave_bias.attr,
 400         NULL
 401 };
 402
 403 static struct attribute_group dbs_attr_group = {
 404         .attrs = dbs_attributes,
 405         .name = "ondemand",
 406 };
 407
 408 /*** delete after deprecation time ***/
 409
 410 #define write_one_old(file_name)                                        \
 411 static ssize_t store_##file_name##_old                                  \
 412 (struct cpufreq_policy *unused, const char *buf, size_t count)          \
 413 {                                                                       \
 414        printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "        \
 415                    "interface is deprecated - " #file_name "\n");       \
 416        return store_##file_name(NULL, NULL, buf, count);                \
 417 }
 418 write_one_old(sampling_rate);
 419 write_one_old(up_threshold);
 420 write_one_old(ignore_nice_load);
 421 write_one_old(powersave_bias);
 422
 423 #define define_one_rw_old(object, _name)       \
 424 static struct freq_attr object =               \
 425 __ATTR(_name, 0644, show_##_name##_old, store_##_name##_old)
 426
 427 define_one_rw_old(sampling_rate_old, sampling_rate);
 428 define_one_rw_old(up_threshold_old, up_threshold);
 429 define_one_rw_old(ignore_nice_load_old, ignore_nice_load);
 430 define_one_rw_old(powersave_bias_old, powersave_bias);
 431
 432 static struct attribute *dbs_attributes_old[] = {
 433        &sampling_rate_max_old.attr,
 434        &sampling_rate_min_old.attr,
 435        &sampling_rate_old.attr,
 436        &up_threshold_old.attr,
 437        &ignore_nice_load_old.attr,
 438        &powersave_bias_old.attr,
 439        NULL
 440 };
 441
 442 static struct attribute_group dbs_attr_group_old = {
 443        .attrs = dbs_attributes_old,
 444        .name = "ondemand",
 445 };
 446
 447 /*** delete after deprecation time ***/
 448
 449 /************************** sysfs end ************************/
 450
 451 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
 452 {
 453         unsigned int max_load_freq;
 454
 455         struct cpufreq_policy *policy;
 456         unsigned int j;
 457
 458         this_dbs_info->freq_lo = 0;
 459         policy = this_dbs_info->cur_policy;
 460
 461         /*
 462          * Every sampling_rate, we check, if current idle time is less
 463          * than 20% (default), then we try to increase frequency
 464          * Every sampling_rate, we look for a the lowest
 465          * frequency which can sustain the load while keeping idle time over
 466          * 30%. If such a frequency exist, we try to decrease to this frequency.
 467          *
 468          * Any frequency increase takes it to the maximum frequency.
 469          * Frequency reduction happens at minimum steps of
 470          * 5% (default) of current frequency
 471          */
 472
 473         /* Get Absolute Load - in terms of freq */
 474         max_load_freq = 0;
 475
 476         for_each_cpu(j, policy->cpus) {
 477                 struct cpu_dbs_info_s *j_dbs_info;
 478                 cputime64_t cur_wall_time, cur_idle_time;
 479                 unsigned int idle_time, wall_time;
 480                 unsigned int load, load_freq;
 481                 int freq_avg;
 482
 483                 j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
 484
 485                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
 486
 487                 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
 488                                 j_dbs_info->prev_cpu_wall);
 489                 j_dbs_info->prev_cpu_wall = cur_wall_time;
 490
 491                 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
 492                                 j_dbs_info->prev_cpu_idle);
 493                 j_dbs_info->prev_cpu_idle = cur_idle_time;
 494
 495                 if (dbs_tuners_ins.ignore_nice) {
 496                         cputime64_t cur_nice;
 497                         unsigned long cur_nice_jiffies;
 498
 499                         cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
 500                                          j_dbs_info->prev_cpu_nice);
 501                         /*
 502                          * Assumption: nice time between sampling periods will
 503                          * be less than 2^32 jiffies for 32 bit sys
 504                          */
 505                         cur_nice_jiffies = (unsigned long)
 506                                         cputime64_to_jiffies64(cur_nice);
 507
 508                         j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 509                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
 510                 }
 511
 512                 if (unlikely(!wall_time || wall_time < idle_time))
 513                         continue;
 514
 515                 load = 100 * (wall_time - idle_time) / wall_time;
 516
 517                 freq_avg = __cpufreq_driver_getavg(policy, j);
 518                 if (freq_avg <= 0)
 519                         freq_avg = policy->cur;
 520
 521                 load_freq = load * freq_avg;
 522                 if (load_freq > max_load_freq)
 523                         max_load_freq = load_freq;
 524         }
 525
 526         /* Check for frequency increase */
 527         if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
 528                 /* if we are already at full speed then break out early */
 529                 if (!dbs_tuners_ins.powersave_bias) {
 530                         if (policy->cur == policy->max)
 531                                 return;
 532
 533                         __cpufreq_driver_target(policy, policy->max,
 534                                 CPUFREQ_RELATION_H);
 535                 } else {
 536                         int freq = powersave_bias_target(policy, policy->max,
 537                                         CPUFREQ_RELATION_H);
 538                         __cpufreq_driver_target(policy, freq,
 539                                 CPUFREQ_RELATION_L);
 540                 }
 541                 return;
 542         }
 543
 544         /* Check for frequency decrease */
 545         /* if we cannot reduce the frequency anymore, break out early */
 546         if (policy->cur == policy->min)
 547                 return;
 548
 549         /*
 550          * The optimal frequency is the frequency that is the lowest that
 551          * can support the current CPU usage without triggering the up
 552          * policy. To be safe, we focus 10 points under the threshold.
 553          */
 554         if (max_load_freq <
 555             (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
 556              policy->cur) {
 557                 unsigned int freq_next;
 558                 freq_next = max_load_freq /
 559                                 (dbs_tuners_ins.up_threshold -
 560                                  dbs_tuners_ins.down_differential);
 561
 562                 if (!dbs_tuners_ins.powersave_bias) {
 563                         __cpufreq_driver_target(policy, freq_next,
 564                                         CPUFREQ_RELATION_L);
 565                 } else {
 566                         int freq = powersave_bias_target(policy, freq_next,
 567                                         CPUFREQ_RELATION_L);
 568                         __cpufreq_driver_target(policy, freq,
 569                                 CPUFREQ_RELATION_L);
 570                 }
 571         }
 572 }
 573
 574 static void do_dbs_timer(struct work_struct *work)
 575 {
 576         struct cpu_dbs_info_s *dbs_info =
 577                 container_of(work, struct cpu_dbs_info_s, work.work);
 578         unsigned int cpu = dbs_info->cpu;
 579         int sample_type = dbs_info->sample_type;
 580
 581         /* We want all CPUs to do sampling nearly on same jiffy */
 582         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 583
 584         if (num_online_cpus() > 1)
 585                 delay -= jiffies % delay;
 586
 587         mutex_lock(&dbs_info->timer_mutex);
 588
 589         /* Common NORMAL_SAMPLE setup */
 590         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 591         if (!dbs_tuners_ins.powersave_bias ||
 592             sample_type == DBS_NORMAL_SAMPLE) {
 593                 dbs_check_cpu(dbs_info);
 594                 if (dbs_info->freq_lo) {
 595                         /* Setup timer for SUB_SAMPLE */
 596                         dbs_info->sample_type = DBS_SUB_SAMPLE;
 597                         delay = dbs_info->freq_hi_jiffies;
 598                 }
 599         } else {
 600                 __cpufreq_driver_target(dbs_info->cur_policy,
 601                         dbs_info->freq_lo, CPUFREQ_RELATION_H);
 602         }
 603         queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
 604         mutex_unlock(&dbs_info->timer_mutex);
 605 }
 606
 607 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
 608 {
 609         /* We want all CPUs to do sampling nearly on same jiffy */
 610         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 611         delay -= jiffies % delay;
 612
 613         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 614         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
 615         queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work,
 616                 delay);
 617 }
 618
 619 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
 620 {
 621         cancel_delayed_work_sync(&dbs_info->work);
 622 }
 623
 624 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 625                                    unsigned int event)
 626 {
 627         unsigned int cpu = policy->cpu;
 628         struct cpu_dbs_info_s *this_dbs_info;
 629         unsigned int j;
 630         int rc;
 631
 632         this_dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 633
 634         switch (event) {
 635         case CPUFREQ_GOV_START:
 636                 if ((!cpu_online(cpu)) || (!policy->cur))
 637                         return -EINVAL;
 638
 639                 mutex_lock(&dbs_mutex);
 640
 641                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group_old);
 642                 if (rc) {
 643                         mutex_unlock(&dbs_mutex);
 644                         return rc;
 645                 }
 646
 647                 dbs_enable++;
 648                 for_each_cpu(j, policy->cpus) {
 649                         struct cpu_dbs_info_s *j_dbs_info;
 650                         j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
 651                         j_dbs_info->cur_policy = policy;
 652
 653                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 654                                                 &j_dbs_info->prev_cpu_wall);
 655                         if (dbs_tuners_ins.ignore_nice) {
 656                                 j_dbs_info->prev_cpu_nice =
 657                                                 kstat_cpu(j).cpustat.nice;
 658                         }
 659                 }
 660                 this_dbs_info->cpu = cpu;
 661                 ondemand_powersave_bias_init_cpu(cpu);
 662                 /*
 663                  * Start the timerschedule work, when this governor
 664                  * is used for first time
 665                  */
 666                 if (dbs_enable == 1) {
 667                         unsigned int latency;
 668
 669                         rc = sysfs_create_group(cpufreq_global_kobject,
 670                                                 &dbs_attr_group);
 671                         if (rc) {
 672                                 mutex_unlock(&dbs_mutex);
 673                                 return rc;
 674                         }
 675
 676                         /* policy latency is in nS. Convert it to uS first */
 677                         latency = policy->cpuinfo.transition_latency / 1000;
 678                         if (latency == 0)
 679                                 latency = 1;
 680                         /* Bring kernel and HW constraints together */
 681                         min_sampling_rate = max(min_sampling_rate,
 682                                         MIN_LATENCY_MULTIPLIER * latency);
 683                         dbs_tuners_ins.sampling_rate =
 684                                 max(min_sampling_rate,
 685                                     latency * LATENCY_MULTIPLIER);
 686                 }
 687                 mutex_unlock(&dbs_mutex);
 688
 689                 mutex_init(&this_dbs_info->timer_mutex);
 690                 dbs_timer_init(this_dbs_info);
 691                 break;
 692
 693         case CPUFREQ_GOV_STOP:
 694                 dbs_timer_exit(this_dbs_info);
 695
 696                 mutex_lock(&dbs_mutex);
 697                 sysfs_remove_group(&policy->kobj, &dbs_attr_group_old);
 698                 mutex_destroy(&this_dbs_info->timer_mutex);
 699                 dbs_enable--;
 700                 mutex_unlock(&dbs_mutex);
 701                 if (!dbs_enable)
 702                         sysfs_remove_group(cpufreq_global_kobject,
 703                                            &dbs_attr_group);
 704
 705                 break;
 706
 707         case CPUFREQ_GOV_LIMITS:
 708                 mutex_lock(&this_dbs_info->timer_mutex);
 709                 if (policy->max < this_dbs_info->cur_policy->cur)
 710                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 711                                 policy->max, CPUFREQ_RELATION_H);
 712                 else if (policy->min > this_dbs_info->cur_policy->cur)
 713                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 714                                 policy->min, CPUFREQ_RELATION_L);
 715                 mutex_unlock(&this_dbs_info->timer_mutex);
 716                 break;
 717         }
 718         return 0;
 719 }
 720
 721 static int __init cpufreq_gov_dbs_init(void)
 722 {
 723         int err;
 724         cputime64_t wall;
 725         u64 idle_time;
 726         int cpu = get_cpu();
 727
 728         idle_time = get_cpu_idle_time_us(cpu, &wall);
 729         put_cpu();
 730         if (idle_time != -1ULL) {
 731                 /* Idle micro accounting is supported. Use finer thresholds */
 732                 dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
 733                 dbs_tuners_ins.down_differential =
 734                                         MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
 735                 /*
 736                  * In no_hz/micro accounting case we set the minimum frequency
 737                  * not depending on HZ, but fixed (very low). The deferred
 738                  * timer might skip some samples if idle/sleeping as needed.
 739                 */
 740                 min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
 741         } else {
 742                 /* For correct statistics, we need 10 ticks for each measure */
 743                 min_sampling_rate =
 744                         MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
 745         }
 746
 747         kondemand_wq = create_workqueue("kondemand");
 748         if (!kondemand_wq) {
 749                 printk(KERN_ERR "Creation of kondemand failed\n");
 750                 return -EFAULT;
 751         }
 752         err = cpufreq_register_governor(&cpufreq_gov_ondemand);
 753         if (err)
 754                 destroy_workqueue(kondemand_wq);
 755
 756         return err;
 757 }
 758
 759 static void __exit cpufreq_gov_dbs_exit(void)
 760 {
 761         cpufreq_unregister_governor(&cpufreq_gov_ondemand);
 762         destroy_workqueue(kondemand_wq);
 763 }
 764
 765
 766 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 767 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 768 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 769         "Low Latency Frequency Transition capable processors");
 770 MODULE_LICENSE("GPL");
 771
 772 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 773 fs_initcall(cpufreq_gov_dbs_init);
 774 #else
 775 module_init(cpufreq_gov_dbs_init);
 776 #endif
 777 module_exit(cpufreq_gov_dbs_exit);