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 DEF_SAMPLING_DOWN_FACTOR                (1)
  34 #define MAX_SAMPLING_DOWN_FACTOR                (100000)
  35 #define MICRO_FREQUENCY_DOWN_DIFFERENTIAL       (3)
  36 #define MICRO_FREQUENCY_UP_THRESHOLD            (95)
  37 #define MICRO_FREQUENCY_MIN_SAMPLE_RATE         (10000)
  38 #define MIN_FREQUENCY_UP_THRESHOLD              (11)
  39 #define MAX_FREQUENCY_UP_THRESHOLD              (100)
  40
  41 /*
  42  * The polling frequency of this governor depends on the capability of
  43  * the processor. Default polling frequency is 1000 times the transition
  44  * latency of the processor. The governor will work on any processor with
  45  * transition latency <= 10mS, using appropriate sampling
  46  * rate.
  47  * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
  48  * this governor will not work.
  49  * All times here are in uS.
  50  */
  51 #define MIN_SAMPLING_RATE_RATIO                 (2)
  52
  53 static unsigned int min_sampling_rate;
  54
  55 #define LATENCY_MULTIPLIER                      (1000)
  56 #define MIN_LATENCY_MULTIPLIER                  (100)
  57 #define TRANSITION_LATENCY_LIMIT                (10 * 1000 * 1000)
  58
  59 static void do_dbs_timer(struct work_struct *work);
  60 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
  61                                 unsigned int event);
  62
  63 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
  64 static
  65 #endif
  66 struct cpufreq_governor cpufreq_gov_ondemand = {
  67        .name                   = "ondemand",
  68        .governor               = cpufreq_governor_dbs,
  69        .max_transition_latency = TRANSITION_LATENCY_LIMIT,
  70        .owner                  = THIS_MODULE,
  71 };
  72
  73 /* Sampling types */
  74 enum {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE};
  75
  76 struct cpu_dbs_info_s {
  77         cputime64_t prev_cpu_idle;
  78         cputime64_t prev_cpu_iowait;
  79         cputime64_t prev_cpu_wall;
  80         cputime64_t prev_cpu_nice;
  81         struct cpufreq_policy *cur_policy;
  82         struct delayed_work work;
  83         struct cpufreq_frequency_table *freq_table;
  84         unsigned int freq_lo;
  85         unsigned int freq_lo_jiffies;
  86         unsigned int freq_hi_jiffies;
  87         unsigned int rate_mult;
  88         int cpu;
  89         unsigned int sample_type:1;
  90         /*
  91          * percpu mutex that serializes governor limit change with
  92          * do_dbs_timer invocation. We do not want do_dbs_timer to run
  93          * when user is changing the governor or limits.
  94          */
  95         struct mutex timer_mutex;
  96 };
  97 static DEFINE_PER_CPU(struct cpu_dbs_info_s, od_cpu_dbs_info);
  98
  99 static unsigned int dbs_enable; /* number of CPUs using this policy */
 100
 101 /*
 102  * dbs_mutex protects dbs_enable in governor start/stop.
 103  */
 104 static DEFINE_MUTEX(dbs_mutex);
 105
 106 static struct dbs_tuners {
 107         unsigned int sampling_rate;
 108         unsigned int up_threshold;
 109         unsigned int down_differential;
 110         unsigned int ignore_nice;
 111         unsigned int sampling_down_factor;
 112         unsigned int powersave_bias;
 113         unsigned int io_is_busy;
 114 } dbs_tuners_ins = {
 115         .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
 116         .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
 117         .down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL,
 118         .ignore_nice = 0,
 119         .powersave_bias = 0,
 120 };
 121
 122 static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
 123                                                         cputime64_t *wall)
 124 {
 125         cputime64_t idle_time;
 126         cputime64_t cur_wall_time;
 127         cputime64_t busy_time;
 128
 129         cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
 130         busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
 131                         kstat_cpu(cpu).cpustat.system);
 132
 133         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
 134         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
 135         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
 136         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice);
 137
 138         idle_time = cputime64_sub(cur_wall_time, busy_time);
 139         if (wall)
 140                 *wall = (cputime64_t)jiffies_to_usecs(cur_wall_time);
 141
 142         return (cputime64_t)jiffies_to_usecs(idle_time);
 143 }
 144
 145 static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
 146 {
 147         u64 idle_time = get_cpu_idle_time_us(cpu, wall);
 148
 149         if (idle_time == -1ULL)
 150                 return get_cpu_idle_time_jiffy(cpu, wall);
 151
 152         return idle_time;
 153 }
 154
 155 static inline cputime64_t get_cpu_iowait_time(unsigned int cpu, cputime64_t *wall)
 156 {
 157         u64 iowait_time = get_cpu_iowait_time_us(cpu, wall);
 158
 159         if (iowait_time == -1ULL)
 160                 return 0;
 161
 162         return iowait_time;
 163 }
 164
 165 /*
 166  * Find right freq to be set now with powersave_bias on.
 167  * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
 168  * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
 169  */
 170 static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
 171                                           unsigned int freq_next,
 172                                           unsigned int relation)
 173 {
 174         unsigned int freq_req, freq_reduc, freq_avg;
 175         unsigned int freq_hi, freq_lo;
 176         unsigned int index = 0;
 177         unsigned int jiffies_total, jiffies_hi, jiffies_lo;
 178         struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
 179                                                    policy->cpu);
 180
 181         if (!dbs_info->freq_table) {
 182                 dbs_info->freq_lo = 0;
 183                 dbs_info->freq_lo_jiffies = 0;
 184                 return freq_next;
 185         }
 186
 187         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
 188                         relation, &index);
 189         freq_req = dbs_info->freq_table[index].frequency;
 190         freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
 191         freq_avg = freq_req - freq_reduc;
 192
 193         /* Find freq bounds for freq_avg in freq_table */
 194         index = 0;
 195         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 196                         CPUFREQ_RELATION_H, &index);
 197         freq_lo = dbs_info->freq_table[index].frequency;
 198         index = 0;
 199         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 200                         CPUFREQ_RELATION_L, &index);
 201         freq_hi = dbs_info->freq_table[index].frequency;
 202
 203         /* Find out how long we have to be in hi and lo freqs */
 204         if (freq_hi == freq_lo) {
 205                 dbs_info->freq_lo = 0;
 206                 dbs_info->freq_lo_jiffies = 0;
 207                 return freq_lo;
 208         }
 209         jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 210         jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
 211         jiffies_hi += ((freq_hi - freq_lo) / 2);
 212         jiffies_hi /= (freq_hi - freq_lo);
 213         jiffies_lo = jiffies_total - jiffies_hi;
 214         dbs_info->freq_lo = freq_lo;
 215         dbs_info->freq_lo_jiffies = jiffies_lo;
 216         dbs_info->freq_hi_jiffies = jiffies_hi;
 217         return freq_hi;
 218 }
 219
 220 static void ondemand_powersave_bias_init_cpu(int cpu)
 221 {
 222         struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 223         dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
 224         dbs_info->freq_lo = 0;
 225 }
 226
 227 static void ondemand_powersave_bias_init(void)
 228 {
 229         int i;
 230         for_each_online_cpu(i) {
 231                 ondemand_powersave_bias_init_cpu(i);
 232         }
 233 }
 234
 235 /************************** sysfs interface ************************/
 236
 237 static ssize_t show_sampling_rate_min(struct kobject *kobj,
 238                                       struct attribute *attr, char *buf)
 239 {
 240         return sprintf(buf, "%u\n", min_sampling_rate);
 241 }
 242
 243 define_one_global_ro(sampling_rate_min);
 244
 245 /* cpufreq_ondemand Governor Tunables */
 246 #define show_one(file_name, object)                                     \
 247 static ssize_t show_##file_name                                         \
 248 (struct kobject *kobj, struct attribute *attr, char *buf)              \
 249 {                                                                       \
 250         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
 251 }
 252 show_one(sampling_rate, sampling_rate);
 253 show_one(io_is_busy, io_is_busy);
 254 show_one(up_threshold, up_threshold);
 255 show_one(sampling_down_factor, sampling_down_factor);
 256 show_one(ignore_nice_load, ignore_nice);
 257 show_one(powersave_bias, powersave_bias);
 258
 259 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
 260                                    const char *buf, size_t count)
 261 {
 262         unsigned int input;
 263         int ret;
 264         ret = sscanf(buf, "%u", &input);
 265         if (ret != 1)
 266                 return -EINVAL;
 267         dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
 268         return count;
 269 }
 270
 271 static ssize_t store_io_is_busy(struct kobject *a, struct attribute *b,
 272                                    const char *buf, size_t count)
 273 {
 274         unsigned int input;
 275         int ret;
 276
 277         ret = sscanf(buf, "%u", &input);
 278         if (ret != 1)
 279                 return -EINVAL;
 280         dbs_tuners_ins.io_is_busy = !!input;
 281         return count;
 282 }
 283
 284 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
 285                                   const char *buf, size_t count)
 286 {
 287         unsigned int input;
 288         int ret;
 289         ret = sscanf(buf, "%u", &input);
 290
 291         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 292                         input < MIN_FREQUENCY_UP_THRESHOLD) {
 293                 return -EINVAL;
 294         }
 295         dbs_tuners_ins.up_threshold = input;
 296         return count;
 297 }
 298
 299 static ssize_t store_sampling_down_factor(struct kobject *a,
 300                         struct attribute *b, const char *buf, size_t count)
 301 {
 302         unsigned int input, j;
 303         int ret;
 304         ret = sscanf(buf, "%u", &input);
 305
 306         if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
 307                 return -EINVAL;
 308         dbs_tuners_ins.sampling_down_factor = input;
 309
 310         /* Reset down sampling multiplier in case it was active */
 311         for_each_online_cpu(j) {
 312                 struct cpu_dbs_info_s *dbs_info;
 313                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
 314                 dbs_info->rate_mult = 1;
 315         }
 316         return count;
 317 }
 318
 319 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
 320                                       const char *buf, size_t count)
 321 {
 322         unsigned int input;
 323         int ret;
 324
 325         unsigned int j;
 326
 327         ret = sscanf(buf, "%u", &input);
 328         if (ret != 1)
 329                 return -EINVAL;
 330
 331         if (input > 1)
 332                 input = 1;
 333
 334         if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
 335                 return count;
 336         }
 337         dbs_tuners_ins.ignore_nice = input;
 338
 339         /* we need to re-evaluate prev_cpu_idle */
 340         for_each_online_cpu(j) {
 341                 struct cpu_dbs_info_s *dbs_info;
 342                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
 343                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 344                                                 &dbs_info->prev_cpu_wall);
 345                 if (dbs_tuners_ins.ignore_nice)
 346                         dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 347
 348         }
 349         return count;
 350 }
 351
 352 static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b,
 353                                     const char *buf, size_t count)
 354 {
 355         unsigned int input;
 356         int ret;
 357         ret = sscanf(buf, "%u", &input);
 358
 359         if (ret != 1)
 360                 return -EINVAL;
 361
 362         if (input > 1000)
 363                 input = 1000;
 364
 365         dbs_tuners_ins.powersave_bias = input;
 366         ondemand_powersave_bias_init();
 367         return count;
 368 }
 369
 370 define_one_global_rw(sampling_rate);
 371 define_one_global_rw(io_is_busy);
 372 define_one_global_rw(up_threshold);
 373 define_one_global_rw(sampling_down_factor);
 374 define_one_global_rw(ignore_nice_load);
 375 define_one_global_rw(powersave_bias);
 376
 377 static struct attribute *dbs_attributes[] = {
 378         &sampling_rate_min.attr,
 379         &sampling_rate.attr,
 380         &up_threshold.attr,
 381         &sampling_down_factor.attr,
 382         &ignore_nice_load.attr,
 383         &powersave_bias.attr,
 384         &io_is_busy.attr,
 385         NULL
 386 };
 387
 388 static struct attribute_group dbs_attr_group = {
 389         .attrs = dbs_attributes,
 390         .name = "ondemand",
 391 };
 392
 393 /************************** sysfs end ************************/
 394
 395 static void dbs_freq_increase(struct cpufreq_policy *p, unsigned int freq)
 396 {
 397         if (dbs_tuners_ins.powersave_bias)
 398                 freq = powersave_bias_target(p, freq, CPUFREQ_RELATION_H);
 399         else if (p->cur == p->max)
 400                 return;
 401
 402         __cpufreq_driver_target(p, freq, dbs_tuners_ins.powersave_bias ?
 403                         CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
 404 }
 405
 406 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
 407 {
 408         unsigned int max_load_freq;
 409
 410         struct cpufreq_policy *policy;
 411         unsigned int j;
 412
 413         this_dbs_info->freq_lo = 0;
 414         policy = this_dbs_info->cur_policy;
 415
 416         /*
 417          * Every sampling_rate, we check, if current idle time is less
 418          * than 20% (default), then we try to increase frequency
 419          * Every sampling_rate, we look for a the lowest
 420          * frequency which can sustain the load while keeping idle time over
 421          * 30%. If such a frequency exist, we try to decrease to this frequency.
 422          *
 423          * Any frequency increase takes it to the maximum frequency.
 424          * Frequency reduction happens at minimum steps of
 425          * 5% (default) of current frequency
 426          */
 427
 428         /* Get Absolute Load - in terms of freq */
 429         max_load_freq = 0;
 430
 431         for_each_cpu(j, policy->cpus) {
 432                 struct cpu_dbs_info_s *j_dbs_info;
 433                 cputime64_t cur_wall_time, cur_idle_time, cur_iowait_time;
 434                 unsigned int idle_time, wall_time, iowait_time;
 435                 unsigned int load, load_freq;
 436                 int freq_avg;
 437
 438                 j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
 439
 440                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
 441                 cur_iowait_time = get_cpu_iowait_time(j, &cur_wall_time);
 442
 443                 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
 444                                 j_dbs_info->prev_cpu_wall);
 445                 j_dbs_info->prev_cpu_wall = cur_wall_time;
 446
 447                 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
 448                                 j_dbs_info->prev_cpu_idle);
 449                 j_dbs_info->prev_cpu_idle = cur_idle_time;
 450
 451                 iowait_time = (unsigned int) cputime64_sub(cur_iowait_time,
 452                                 j_dbs_info->prev_cpu_iowait);
 453                 j_dbs_info->prev_cpu_iowait = cur_iowait_time;
 454
 455                 if (dbs_tuners_ins.ignore_nice) {
 456                         cputime64_t cur_nice;
 457                         unsigned long cur_nice_jiffies;
 458
 459                         cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
 460                                          j_dbs_info->prev_cpu_nice);
 461                         /*
 462                          * Assumption: nice time between sampling periods will
 463                          * be less than 2^32 jiffies for 32 bit sys
 464                          */
 465                         cur_nice_jiffies = (unsigned long)
 466                                         cputime64_to_jiffies64(cur_nice);
 467
 468                         j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 469                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
 470                 }
 471
 472                 /*
 473                  * For the purpose of ondemand, waiting for disk IO is an
 474                  * indication that you're performance critical, and not that
 475                  * the system is actually idle. So subtract the iowait time
 476                  * from the cpu idle time.
 477                  */
 478
 479                 if (dbs_tuners_ins.io_is_busy && idle_time >= iowait_time)
 480                         idle_time -= iowait_time;
 481
 482                 if (unlikely(!wall_time || wall_time < idle_time))
 483                         continue;
 484
 485                 load = 100 * (wall_time - idle_time) / wall_time;
 486
 487                 freq_avg = __cpufreq_driver_getavg(policy, j);
 488                 if (freq_avg <= 0)
 489                         freq_avg = policy->cur;
 490
 491                 load_freq = load * freq_avg;
 492                 if (load_freq > max_load_freq)
 493                         max_load_freq = load_freq;
 494         }
 495
 496         /* Check for frequency increase */
 497         if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
 498                 /* If switching to max speed, apply sampling_down_factor */
 499                 if (policy->cur < policy->max)
 500                         this_dbs_info->rate_mult =
 501                                 dbs_tuners_ins.sampling_down_factor;
 502                 dbs_freq_increase(policy, policy->max);
 503                 return;
 504         }
 505
 506         /* Check for frequency decrease */
 507         /* if we cannot reduce the frequency anymore, break out early */
 508         if (policy->cur == policy->min)
 509                 return;
 510
 511         /*
 512          * The optimal frequency is the frequency that is the lowest that
 513          * can support the current CPU usage without triggering the up
 514          * policy. To be safe, we focus 10 points under the threshold.
 515          */
 516         if (max_load_freq <
 517             (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
 518              policy->cur) {
 519                 unsigned int freq_next;
 520                 freq_next = max_load_freq /
 521                                 (dbs_tuners_ins.up_threshold -
 522                                  dbs_tuners_ins.down_differential);
 523
 524                 /* No longer fully busy, reset rate_mult */
 525                 this_dbs_info->rate_mult = 1;
 526
 527                 if (freq_next < policy->min)
 528                         freq_next = policy->min;
 529
 530                 if (!dbs_tuners_ins.powersave_bias) {
 531                         __cpufreq_driver_target(policy, freq_next,
 532                                         CPUFREQ_RELATION_L);
 533                 } else {
 534                         int freq = powersave_bias_target(policy, freq_next,
 535                                         CPUFREQ_RELATION_L);
 536                         __cpufreq_driver_target(policy, freq,
 537                                 CPUFREQ_RELATION_L);
 538                 }
 539         }
 540 }
 541
 542 static void do_dbs_timer(struct work_struct *work)
 543 {
 544         struct cpu_dbs_info_s *dbs_info =
 545                 container_of(work, struct cpu_dbs_info_s, work.work);
 546         unsigned int cpu = dbs_info->cpu;
 547         int sample_type = dbs_info->sample_type;
 548
 549         int delay;
 550
 551         mutex_lock(&dbs_info->timer_mutex);
 552
 553         /* Common NORMAL_SAMPLE setup */
 554         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 555         if (!dbs_tuners_ins.powersave_bias ||
 556             sample_type == DBS_NORMAL_SAMPLE) {
 557                 dbs_check_cpu(dbs_info);
 558                 if (dbs_info->freq_lo) {
 559                         /* Setup timer for SUB_SAMPLE */
 560                         dbs_info->sample_type = DBS_SUB_SAMPLE;
 561                         delay = dbs_info->freq_hi_jiffies;
 562                 } else {
 563                         /* We want all CPUs to do sampling nearly on
 564                          * same jiffy
 565                          */
 566                         delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate
 567                                 * dbs_info->rate_mult);
 568
 569                         if (num_online_cpus() > 1)
 570                                 delay -= jiffies % delay;
 571                 }
 572         } else {
 573                 __cpufreq_driver_target(dbs_info->cur_policy,
 574                         dbs_info->freq_lo, CPUFREQ_RELATION_H);
 575                 delay = dbs_info->freq_lo_jiffies;
 576         }
 577         schedule_delayed_work_on(cpu, &dbs_info->work, delay);
 578         mutex_unlock(&dbs_info->timer_mutex);
 579 }
 580
 581 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
 582 {
 583         /* We want all CPUs to do sampling nearly on same jiffy */
 584         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 585
 586         if (num_online_cpus() > 1)
 587                 delay -= jiffies % delay;
 588
 589         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 590         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
 591         schedule_delayed_work_on(dbs_info->cpu, &dbs_info->work, delay);
 592 }
 593
 594 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
 595 {
 596         cancel_delayed_work_sync(&dbs_info->work);
 597 }
 598
 599 /*
 600  * Not all CPUs want IO time to be accounted as busy; this dependson how
 601  * efficient idling at a higher frequency/voltage is.
 602  * Pavel Machek says this is not so for various generations of AMD and old
 603  * Intel systems.
 604  * Mike Chan (androidlcom) calis this is also not true for ARM.
 605  * Because of this, whitelist specific known (series) of CPUs by default, and
 606  * leave all others up to the user.
 607  */
 608 static int should_io_be_busy(void)
 609 {
 610 #if defined(CONFIG_X86)
 611         /*
 612          * For Intel, Core 2 (model 15) andl later have an efficient idle.
 613          */
 614         if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
 615             boot_cpu_data.x86 == 6 &&
 616             boot_cpu_data.x86_model >= 15)
 617                 return 1;
 618 #endif
 619         return 0;
 620 }
 621
 622 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 623                                    unsigned int event)
 624 {
 625         unsigned int cpu = policy->cpu;
 626         struct cpu_dbs_info_s *this_dbs_info;
 627         unsigned int j;
 628         int rc;
 629
 630         this_dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 631
 632         switch (event) {
 633         case CPUFREQ_GOV_START:
 634                 if ((!cpu_online(cpu)) || (!policy->cur))
 635                         return -EINVAL;
 636
 637                 mutex_lock(&dbs_mutex);
 638
 639                 dbs_enable++;
 640                 for_each_cpu(j, policy->cpus) {
 641                         struct cpu_dbs_info_s *j_dbs_info;
 642                         j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
 643                         j_dbs_info->cur_policy = policy;
 644
 645                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 646                                                 &j_dbs_info->prev_cpu_wall);
 647                         if (dbs_tuners_ins.ignore_nice) {
 648                                 j_dbs_info->prev_cpu_nice =
 649                                                 kstat_cpu(j).cpustat.nice;
 650                         }
 651                 }
 652                 this_dbs_info->cpu = cpu;
 653                 this_dbs_info->rate_mult = 1;
 654                 ondemand_powersave_bias_init_cpu(cpu);
 655                 /*
 656                  * Start the timerschedule work, when this governor
 657                  * is used for first time
 658                  */
 659                 if (dbs_enable == 1) {
 660                         unsigned int latency;
 661
 662                         rc = sysfs_create_group(cpufreq_global_kobject,
 663                                                 &dbs_attr_group);
 664                         if (rc) {
 665                                 mutex_unlock(&dbs_mutex);
 666                                 return rc;
 667                         }
 668
 669                         /* policy latency is in nS. Convert it to uS first */
 670                         latency = policy->cpuinfo.transition_latency / 1000;
 671                         if (latency == 0)
 672                                 latency = 1;
 673                         /* Bring kernel and HW constraints together */
 674                         min_sampling_rate = max(min_sampling_rate,
 675                                         MIN_LATENCY_MULTIPLIER * latency);
 676                         dbs_tuners_ins.sampling_rate =
 677                                 max(min_sampling_rate,
 678                                     latency * LATENCY_MULTIPLIER);
 679                         dbs_tuners_ins.io_is_busy = should_io_be_busy();
 680                 }
 681                 mutex_unlock(&dbs_mutex);
 682
 683                 mutex_init(&this_dbs_info->timer_mutex);
 684                 dbs_timer_init(this_dbs_info);
 685                 break;
 686
 687         case CPUFREQ_GOV_STOP:
 688                 dbs_timer_exit(this_dbs_info);
 689
 690                 mutex_lock(&dbs_mutex);
 691                 mutex_destroy(&this_dbs_info->timer_mutex);
 692                 dbs_enable--;
 693                 mutex_unlock(&dbs_mutex);
 694                 if (!dbs_enable)
 695                         sysfs_remove_group(cpufreq_global_kobject,
 696                                            &dbs_attr_group);
 697
 698                 break;
 699
 700         case CPUFREQ_GOV_LIMITS:
 701                 mutex_lock(&this_dbs_info->timer_mutex);
 702                 if (policy->max < this_dbs_info->cur_policy->cur)
 703                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 704                                 policy->max, CPUFREQ_RELATION_H);
 705                 else if (policy->min > this_dbs_info->cur_policy->cur)
 706                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 707                                 policy->min, CPUFREQ_RELATION_L);
 708                 mutex_unlock(&this_dbs_info->timer_mutex);
 709                 break;
 710         }
 711         return 0;
 712 }
 713
 714 static int __init cpufreq_gov_dbs_init(void)
 715 {
 716         cputime64_t wall;
 717         u64 idle_time;
 718         int cpu = get_cpu();
 719
 720         idle_time = get_cpu_idle_time_us(cpu, &wall);
 721         put_cpu();
 722         if (idle_time != -1ULL) {
 723                 /* Idle micro accounting is supported. Use finer thresholds */
 724                 dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
 725                 dbs_tuners_ins.down_differential =
 726                                         MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
 727                 /*
 728                  * In nohz/micro accounting case we set the minimum frequency
 729                  * not depending on HZ, but fixed (very low). The deferred
 730                  * timer might skip some samples if idle/sleeping as needed.
 731                 */
 732                 min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
 733         } else {
 734                 /* For correct statistics, we need 10 ticks for each measure */
 735                 min_sampling_rate =
 736                         MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
 737         }
 738
 739         return cpufreq_register_governor(&cpufreq_gov_ondemand);
 740 }
 741
 742 static void __exit cpufreq_gov_dbs_exit(void)
 743 {
 744         cpufreq_unregister_governor(&cpufreq_gov_ondemand);
 745 }
 746
 747
 748 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 749 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 750 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 751         "Low Latency Frequency Transition capable processors");
 752 MODULE_LICENSE("GPL");
 753
 754 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 755 fs_initcall(cpufreq_gov_dbs_init);
 756 #else
 757 module_init(cpufreq_gov_dbs_init);
 758 #endif
 759 module_exit(cpufreq_gov_dbs_exit);