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 #define MAX_DEFAULT_SAMPLING_RATE               (300 * 1000U)
  41
  42 /*
  43  * The polling frequency of this governor depends on the capability of
  44  * the processor. Default polling frequency is 1000 times the transition
  45  * latency of the processor. The governor will work on any processor with
  46  * transition latency <= 10mS, using appropriate sampling
  47  * rate.
  48  * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
  49  * this governor will not work.
  50  * All times here are in uS.
  51  */
  52 #define MIN_SAMPLING_RATE_RATIO                 (2)
  53
  54 static unsigned int min_sampling_rate;
  55
  56 #define LATENCY_MULTIPLIER                      (1000)
  57 #define MIN_LATENCY_MULTIPLIER                  (100)
  58 #define TRANSITION_LATENCY_LIMIT                (10 * 1000 * 1000)
  59
  60 static void do_dbs_timer(struct work_struct *work);
  61 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
  62                                 unsigned int event);
  63
  64 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
  65 static
  66 #endif
  67 struct cpufreq_governor cpufreq_gov_ondemand = {
  68        .name                   = "ondemand",
  69        .governor               = cpufreq_governor_dbs,
  70        .max_transition_latency = TRANSITION_LATENCY_LIMIT,
  71        .owner                  = THIS_MODULE,
  72 };
  73
  74 /* Sampling types */
  75 enum {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE};
  76
  77 struct cpu_dbs_info_s {
  78         cputime64_t prev_cpu_idle;
  79         cputime64_t prev_cpu_iowait;
  80         cputime64_t prev_cpu_wall;
  81         cputime64_t prev_cpu_nice;
  82         struct cpufreq_policy *cur_policy;
  83         struct delayed_work work;
  84         struct cpufreq_frequency_table *freq_table;
  85         unsigned int freq_lo;
  86         unsigned int freq_lo_jiffies;
  87         unsigned int freq_hi_jiffies;
  88         unsigned int rate_mult;
  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 dbs_enable in governor start/stop.
 104  */
 105 static DEFINE_MUTEX(dbs_mutex);
 106
 107 static struct dbs_tuners {
 108         unsigned int sampling_rate;
 109         unsigned int up_threshold;
 110         unsigned int down_differential;
 111         unsigned int ignore_nice;
 112         unsigned int sampling_down_factor;
 113         unsigned int powersave_bias;
 114         unsigned int io_is_busy;
 115 } dbs_tuners_ins = {
 116         .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
 117         .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
 118         .down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL,
 119         .ignore_nice = 0,
 120         .powersave_bias = 0,
 121 };
 122
 123 static inline u64 get_cpu_idle_time_jiffy(unsigned int cpu, u64 *wall)
 124 {
 125         u64 idle_time;
 126         u64 cur_wall_time;
 127         u64 busy_time;
 128
 129         cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
 130
 131         busy_time  = kcpustat_cpu(cpu).cpustat[CPUTIME_USER];
 132         busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_SYSTEM];
 133         busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_IRQ];
 134         busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_SOFTIRQ];
 135         busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_STEAL];
 136         busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_NICE];
 137
 138         idle_time = cur_wall_time - busy_time;
 139         if (wall)
 140                 *wall = jiffies_to_usecs(cur_wall_time);
 141
 142         return 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, NULL);
 148
 149         if (idle_time == -1ULL)
 150                 return get_cpu_idle_time_jiffy(cpu, wall);
 151         else
 152                 idle_time += get_cpu_iowait_time_us(cpu, wall);
 153
 154         return idle_time;
 155 }
 156
 157 static inline cputime64_t get_cpu_iowait_time(unsigned int cpu, cputime64_t *wall)
 158 {
 159         u64 iowait_time = get_cpu_iowait_time_us(cpu, wall);
 160
 161         if (iowait_time == -1ULL)
 162                 return 0;
 163
 164         return iowait_time;
 165 }
 166
 167 /*
 168  * Find right freq to be set now with powersave_bias on.
 169  * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
 170  * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
 171  */
 172 static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
 173                                           unsigned int freq_next,
 174                                           unsigned int relation)
 175 {
 176         unsigned int freq_req, freq_reduc, freq_avg;
 177         unsigned int freq_hi, freq_lo;
 178         unsigned int index = 0;
 179         unsigned int jiffies_total, jiffies_hi, jiffies_lo;
 180         struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
 181                                                    policy->cpu);
 182
 183         if (!dbs_info->freq_table) {
 184                 dbs_info->freq_lo = 0;
 185                 dbs_info->freq_lo_jiffies = 0;
 186                 return freq_next;
 187         }
 188
 189         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
 190                         relation, &index);
 191         freq_req = dbs_info->freq_table[index].frequency;
 192         freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
 193         freq_avg = freq_req - freq_reduc;
 194
 195         /* Find freq bounds for freq_avg in freq_table */
 196         index = 0;
 197         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 198                         CPUFREQ_RELATION_H, &index);
 199         freq_lo = dbs_info->freq_table[index].frequency;
 200         index = 0;
 201         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 202                         CPUFREQ_RELATION_L, &index);
 203         freq_hi = dbs_info->freq_table[index].frequency;
 204
 205         /* Find out how long we have to be in hi and lo freqs */
 206         if (freq_hi == freq_lo) {
 207                 dbs_info->freq_lo = 0;
 208                 dbs_info->freq_lo_jiffies = 0;
 209                 return freq_lo;
 210         }
 211         jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 212         jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
 213         jiffies_hi += ((freq_hi - freq_lo) / 2);
 214         jiffies_hi /= (freq_hi - freq_lo);
 215         jiffies_lo = jiffies_total - jiffies_hi;
 216         dbs_info->freq_lo = freq_lo;
 217         dbs_info->freq_lo_jiffies = jiffies_lo;
 218         dbs_info->freq_hi_jiffies = jiffies_hi;
 219         return freq_hi;
 220 }
 221
 222 static void ondemand_powersave_bias_init_cpu(int cpu)
 223 {
 224         struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 225         dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
 226         dbs_info->freq_lo = 0;
 227 }
 228
 229 static void ondemand_powersave_bias_init(void)
 230 {
 231         int i;
 232         for_each_online_cpu(i) {
 233                 ondemand_powersave_bias_init_cpu(i);
 234         }
 235 }
 236
 237 /************************** sysfs interface ************************/
 238
 239 static ssize_t show_sampling_rate_min(struct kobject *kobj,
 240                                       struct attribute *attr, char *buf)
 241 {
 242         return sprintf(buf, "%u\n", min_sampling_rate);
 243 }
 244
 245 define_one_global_ro(sampling_rate_min);
 246
 247 /* cpufreq_ondemand Governor Tunables */
 248 #define show_one(file_name, object)                                     \
 249 static ssize_t show_##file_name                                         \
 250 (struct kobject *kobj, struct attribute *attr, char *buf)              \
 251 {                                                                       \
 252         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
 253 }
 254 show_one(sampling_rate, sampling_rate);
 255 show_one(io_is_busy, io_is_busy);
 256 show_one(up_threshold, up_threshold);
 257 show_one(sampling_down_factor, sampling_down_factor);
 258 show_one(ignore_nice_load, ignore_nice);
 259 show_one(powersave_bias, powersave_bias);
 260
 261 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
 262                                    const char *buf, size_t count)
 263 {
 264         unsigned int input;
 265         int ret;
 266         ret = sscanf(buf, "%u", &input);
 267         if (ret != 1)
 268                 return -EINVAL;
 269         dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
 270         return count;
 271 }
 272
 273 static ssize_t store_io_is_busy(struct kobject *a, struct attribute *b,
 274                                    const char *buf, size_t count)
 275 {
 276         unsigned int input;
 277         int ret;
 278
 279         ret = sscanf(buf, "%u", &input);
 280         if (ret != 1)
 281                 return -EINVAL;
 282         dbs_tuners_ins.io_is_busy = !!input;
 283         return count;
 284 }
 285
 286 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
 287                                   const char *buf, size_t count)
 288 {
 289         unsigned int input;
 290         int ret;
 291         ret = sscanf(buf, "%u", &input);
 292
 293         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 294                         input < MIN_FREQUENCY_UP_THRESHOLD) {
 295                 return -EINVAL;
 296         }
 297         dbs_tuners_ins.up_threshold = input;
 298         return count;
 299 }
 300
 301 static ssize_t store_sampling_down_factor(struct kobject *a,
 302                         struct attribute *b, const char *buf, size_t count)
 303 {
 304         unsigned int input, j;
 305         int ret;
 306         ret = sscanf(buf, "%u", &input);
 307
 308         if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
 309                 return -EINVAL;
 310         dbs_tuners_ins.sampling_down_factor = input;
 311
 312         /* Reset down sampling multiplier in case it was active */
 313         for_each_online_cpu(j) {
 314                 struct cpu_dbs_info_s *dbs_info;
 315                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
 316                 dbs_info->rate_mult = 1;
 317         }
 318         return count;
 319 }
 320
 321 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
 322                                       const char *buf, size_t count)
 323 {
 324         unsigned int input;
 325         int ret;
 326
 327         unsigned int j;
 328
 329         ret = sscanf(buf, "%u", &input);
 330         if (ret != 1)
 331                 return -EINVAL;
 332
 333         if (input > 1)
 334                 input = 1;
 335
 336         if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
 337                 return count;
 338         }
 339         dbs_tuners_ins.ignore_nice = input;
 340
 341         /* we need to re-evaluate prev_cpu_idle */
 342         for_each_online_cpu(j) {
 343                 struct cpu_dbs_info_s *dbs_info;
 344                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
 345                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 346                                                 &dbs_info->prev_cpu_wall);
 347                 if (dbs_tuners_ins.ignore_nice)
 348                         dbs_info->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 349
 350         }
 351         return count;
 352 }
 353
 354 static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b,
 355                                     const char *buf, size_t count)
 356 {
 357         unsigned int input;
 358         int ret;
 359         ret = sscanf(buf, "%u", &input);
 360
 361         if (ret != 1)
 362                 return -EINVAL;
 363
 364         if (input > 1000)
 365                 input = 1000;
 366
 367         dbs_tuners_ins.powersave_bias = input;
 368         ondemand_powersave_bias_init();
 369         return count;
 370 }
 371
 372 define_one_global_rw(sampling_rate);
 373 define_one_global_rw(io_is_busy);
 374 define_one_global_rw(up_threshold);
 375 define_one_global_rw(sampling_down_factor);
 376 define_one_global_rw(ignore_nice_load);
 377 define_one_global_rw(powersave_bias);
 378
 379 static struct attribute *dbs_attributes[] = {
 380         &sampling_rate_min.attr,
 381         &sampling_rate.attr,
 382         &up_threshold.attr,
 383         &sampling_down_factor.attr,
 384         &ignore_nice_load.attr,
 385         &powersave_bias.attr,
 386         &io_is_busy.attr,
 387         NULL
 388 };
 389
 390 static struct attribute_group dbs_attr_group = {
 391         .attrs = dbs_attributes,
 392         .name = "ondemand",
 393 };
 394
 395 /************************** sysfs end ************************/
 396
 397 static void dbs_freq_increase(struct cpufreq_policy *p, unsigned int freq)
 398 {
 399         if (dbs_tuners_ins.powersave_bias)
 400                 freq = powersave_bias_target(p, freq, CPUFREQ_RELATION_H);
 401         else if (p->cur == p->max)
 402                 return;
 403
 404         __cpufreq_driver_target(p, freq, dbs_tuners_ins.powersave_bias ?
 405                         CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
 406 }
 407
 408 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
 409 {
 410         unsigned int max_load_freq;
 411
 412         struct cpufreq_policy *policy;
 413         unsigned int j;
 414
 415         this_dbs_info->freq_lo = 0;
 416         policy = this_dbs_info->cur_policy;
 417
 418         /*
 419          * Every sampling_rate, we check, if current idle time is less
 420          * than 20% (default), then we try to increase frequency
 421          * Every sampling_rate, we look for a the lowest
 422          * frequency which can sustain the load while keeping idle time over
 423          * 30%. If such a frequency exist, we try to decrease to this frequency.
 424          *
 425          * Any frequency increase takes it to the maximum frequency.
 426          * Frequency reduction happens at minimum steps of
 427          * 5% (default) of current frequency
 428          */
 429
 430         /* Get Absolute Load - in terms of freq */
 431         max_load_freq = 0;
 432
 433         for_each_cpu(j, policy->cpus) {
 434                 struct cpu_dbs_info_s *j_dbs_info;
 435                 cputime64_t cur_wall_time, cur_idle_time, cur_iowait_time;
 436                 unsigned int idle_time, wall_time, iowait_time;
 437                 unsigned int load, load_freq;
 438                 int freq_avg;
 439
 440                 j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
 441
 442                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
 443                 cur_iowait_time = get_cpu_iowait_time(j, &cur_wall_time);
 444
 445                 wall_time = (unsigned int)
 446                         (cur_wall_time - j_dbs_info->prev_cpu_wall);
 447                 j_dbs_info->prev_cpu_wall = cur_wall_time;
 448
 449                 idle_time = (unsigned int)
 450                         (cur_idle_time - j_dbs_info->prev_cpu_idle);
 451                 j_dbs_info->prev_cpu_idle = cur_idle_time;
 452
 453                 iowait_time = (unsigned int)
 454                         (cur_iowait_time - j_dbs_info->prev_cpu_iowait);
 455                 j_dbs_info->prev_cpu_iowait = cur_iowait_time;
 456
 457                 if (dbs_tuners_ins.ignore_nice) {
 458                         u64 cur_nice;
 459                         unsigned long cur_nice_jiffies;
 460
 461                         cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] -
 462                                          j_dbs_info->prev_cpu_nice;
 463                         /*
 464                          * Assumption: nice time between sampling periods will
 465                          * be less than 2^32 jiffies for 32 bit sys
 466                          */
 467                         cur_nice_jiffies = (unsigned long)
 468                                         cputime64_to_jiffies64(cur_nice);
 469
 470                         j_dbs_info->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 471                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
 472                 }
 473
 474                 /*
 475                  * For the purpose of ondemand, waiting for disk IO is an
 476                  * indication that you're performance critical, and not that
 477                  * the system is actually idle. So subtract the iowait time
 478                  * from the cpu idle time.
 479                  */
 480
 481                 if (dbs_tuners_ins.io_is_busy && idle_time >= iowait_time)
 482                         idle_time -= iowait_time;
 483
 484                 if (unlikely(!wall_time || wall_time < idle_time))
 485                         continue;
 486
 487                 load = 100 * (wall_time - idle_time) / wall_time;
 488
 489                 freq_avg = __cpufreq_driver_getavg(policy, j);
 490                 if (freq_avg <= 0)
 491                         freq_avg = policy->cur;
 492
 493                 load_freq = load * freq_avg;
 494                 if (load_freq > max_load_freq)
 495                         max_load_freq = load_freq;
 496         }
 497
 498         /* Check for frequency increase */
 499         if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
 500                 /* If switching to max speed, apply sampling_down_factor */
 501                 if (policy->cur < policy->max)
 502                         this_dbs_info->rate_mult =
 503                                 dbs_tuners_ins.sampling_down_factor;
 504                 dbs_freq_increase(policy, policy->max);
 505                 return;
 506         }
 507
 508         /* Check for frequency decrease */
 509         /* if we cannot reduce the frequency anymore, break out early */
 510         if (policy->cur == policy->min)
 511                 return;
 512
 513         /*
 514          * The optimal frequency is the frequency that is the lowest that
 515          * can support the current CPU usage without triggering the up
 516          * policy. To be safe, we focus 10 points under the threshold.
 517          */
 518         if (max_load_freq <
 519             (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
 520              policy->cur) {
 521                 unsigned int freq_next;
 522                 freq_next = max_load_freq /
 523                                 (dbs_tuners_ins.up_threshold -
 524                                  dbs_tuners_ins.down_differential);
 525
 526                 /* No longer fully busy, reset rate_mult */
 527                 this_dbs_info->rate_mult = 1;
 528
 529                 if (freq_next < policy->min)
 530                         freq_next = policy->min;
 531
 532                 if (!dbs_tuners_ins.powersave_bias) {
 533                         __cpufreq_driver_target(policy, freq_next,
 534                                         CPUFREQ_RELATION_L);
 535                 } else {
 536                         int freq = powersave_bias_target(policy, freq_next,
 537                                         CPUFREQ_RELATION_L);
 538                         __cpufreq_driver_target(policy, freq,
 539                                 CPUFREQ_RELATION_L);
 540                 }
 541         }
 542 }
 543
 544 static void do_dbs_timer(struct work_struct *work)
 545 {
 546         struct cpu_dbs_info_s *dbs_info =
 547                 container_of(work, struct cpu_dbs_info_s, work.work);
 548         unsigned int cpu = dbs_info->cpu;
 549         int sample_type = dbs_info->sample_type;
 550
 551         int delay;
 552
 553         mutex_lock(&dbs_info->timer_mutex);
 554
 555         /* Common NORMAL_SAMPLE setup */
 556         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 557         if (!dbs_tuners_ins.powersave_bias ||
 558             sample_type == DBS_NORMAL_SAMPLE) {
 559                 dbs_check_cpu(dbs_info);
 560                 if (dbs_info->freq_lo) {
 561                         /* Setup timer for SUB_SAMPLE */
 562                         dbs_info->sample_type = DBS_SUB_SAMPLE;
 563                         delay = dbs_info->freq_hi_jiffies;
 564                 } else {
 565                         /* We want all CPUs to do sampling nearly on
 566                          * same jiffy
 567                          */
 568                         delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate
 569                                 * dbs_info->rate_mult);
 570
 571                         if (num_online_cpus() > 1)
 572                                 delay -= jiffies % delay;
 573                 }
 574         } else {
 575                 __cpufreq_driver_target(dbs_info->cur_policy,
 576                         dbs_info->freq_lo, CPUFREQ_RELATION_H);
 577                 delay = dbs_info->freq_lo_jiffies;
 578         }
 579         schedule_delayed_work_on(cpu, &dbs_info->work, delay);
 580         mutex_unlock(&dbs_info->timer_mutex);
 581 }
 582
 583 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
 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         if (num_online_cpus() > 1)
 589                 delay -= jiffies % delay;
 590
 591         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 592         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
 593         schedule_delayed_work_on(dbs_info->cpu, &dbs_info->work, delay);
 594 }
 595
 596 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
 597 {
 598         cancel_delayed_work_sync(&dbs_info->work);
 599 }
 600
 601 /*
 602  * Not all CPUs want IO time to be accounted as busy; this dependson how
 603  * efficient idling at a higher frequency/voltage is.
 604  * Pavel Machek says this is not so for various generations of AMD and old
 605  * Intel systems.
 606  * Mike Chan (androidlcom) calis this is also not true for ARM.
 607  * Because of this, whitelist specific known (series) of CPUs by default, and
 608  * leave all others up to the user.
 609  */
 610 static int should_io_be_busy(void)
 611 {
 612 #if defined(CONFIG_X86)
 613         /*
 614          * For Intel, Core 2 (model 15) andl later have an efficient idle.
 615          */
 616         if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
 617             boot_cpu_data.x86 == 6 &&
 618             boot_cpu_data.x86_model >= 15)
 619                 return 1;
 620 #endif
 621         return 0;
 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                 dbs_enable++;
 642                 for_each_cpu(j, policy->cpus) {
 643                         struct cpu_dbs_info_s *j_dbs_info;
 644                         j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
 645                         j_dbs_info->cur_policy = policy;
 646
 647                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 648                                                 &j_dbs_info->prev_cpu_wall);
 649                         if (dbs_tuners_ins.ignore_nice)
 650                                 j_dbs_info->prev_cpu_nice =
 651                                                 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 652                 }
 653                 this_dbs_info->cpu = cpu;
 654                 this_dbs_info->rate_mult = 1;
 655                 ondemand_powersave_bias_init_cpu(cpu);
 656                 /*
 657                  * Start the timerschedule work, when this governor
 658                  * is used for first time
 659                  */
 660                 if (dbs_enable == 1) {
 661                         unsigned int latency;
 662
 663                         rc = sysfs_create_group(cpufreq_global_kobject,
 664                                                 &dbs_attr_group);
 665                         if (rc) {
 666                                 mutex_unlock(&dbs_mutex);
 667                                 return rc;
 668                         }
 669
 670                         /* policy latency is in nS. Convert it to uS first */
 671                         latency = policy->cpuinfo.transition_latency / 1000;
 672                         if (latency == 0)
 673                                 latency = 1;
 674                         /* Bring kernel and HW constraints together */
 675                         min_sampling_rate = max(min_sampling_rate,
 676                                         MIN_LATENCY_MULTIPLIER * latency);
 677                         dbs_tuners_ins.sampling_rate =
 678                                 max(min_sampling_rate,
 679                                     latency * LATENCY_MULTIPLIER);
 680                         /*
 681                          * Cut def_sampling rate to 300ms if it was above,
 682                          * still consider to not set it above latency
 683                          * transition * 100
 684                          */
 685                         if (dbs_tuners_ins.sampling_rate > MAX_DEFAULT_SAMPLING_RATE) {
 686                                 dbs_tuners_ins.sampling_rate =
 687                                         max(min_sampling_rate, MAX_DEFAULT_SAMPLING_RATE);
 688                                 printk(KERN_INFO "CPUFREQ: ondemand sampling "
 689                                        "rate set to %d ms\n",
 690                                        dbs_tuners_ins.sampling_rate / 1000);
 691                         }
 692                         /*
 693                          * Be conservative in respect to performance.
 694                          * If an application calculates using two threads
 695                          * depending on each other, they will be run on several
 696                          * CPU cores resulting on 50% load on both.
 697                          * SLED might still want to prefer 80% up_threshold
 698                          * by default, but we cannot differ that here.
 699                          */
 700                         if (num_online_cpus() > 1)
 701                                 dbs_tuners_ins.up_threshold =
 702                                         DEF_FREQUENCY_UP_THRESHOLD / 2;
 703                         dbs_tuners_ins.io_is_busy = should_io_be_busy();
 704                 }
 705                 mutex_unlock(&dbs_mutex);
 706
 707                 mutex_init(&this_dbs_info->timer_mutex);
 708                 dbs_timer_init(this_dbs_info);
 709                 break;
 710
 711         case CPUFREQ_GOV_STOP:
 712                 dbs_timer_exit(this_dbs_info);
 713
 714                 mutex_lock(&dbs_mutex);
 715                 mutex_destroy(&this_dbs_info->timer_mutex);
 716                 dbs_enable--;
 717                 mutex_unlock(&dbs_mutex);
 718                 if (!dbs_enable)
 719                         sysfs_remove_group(cpufreq_global_kobject,
 720                                            &dbs_attr_group);
 721
 722                 break;
 723
 724         case CPUFREQ_GOV_LIMITS:
 725                 mutex_lock(&this_dbs_info->timer_mutex);
 726                 if (policy->max < this_dbs_info->cur_policy->cur)
 727                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 728                                 policy->max, CPUFREQ_RELATION_H);
 729                 else if (policy->min > this_dbs_info->cur_policy->cur)
 730                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 731                                 policy->min, CPUFREQ_RELATION_L);
 732                 mutex_unlock(&this_dbs_info->timer_mutex);
 733                 break;
 734         }
 735         return 0;
 736 }
 737
 738 static int __init cpufreq_gov_dbs_init(void)
 739 {
 740         u64 idle_time;
 741         int cpu = get_cpu();
 742
 743         idle_time = get_cpu_idle_time_us(cpu, NULL);
 744         put_cpu();
 745         if (idle_time != -1ULL) {
 746                 /* Idle micro accounting is supported. Use finer thresholds */
 747                 dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
 748                 dbs_tuners_ins.down_differential =
 749                                         MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
 750                 /*
 751                  * In nohz/micro accounting case we set the minimum frequency
 752                  * not depending on HZ, but fixed (very low). The deferred
 753                  * timer might skip some samples if idle/sleeping as needed.
 754                 */
 755                 min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
 756         } else {
 757                 /* For correct statistics, we need 10 ticks for each measure */
 758                 min_sampling_rate =
 759                         MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
 760         }
 761
 762         return cpufreq_register_governor(&cpufreq_gov_ondemand);
 763 }
 764
 765 static void __exit cpufreq_gov_dbs_exit(void)
 766 {
 767         cpufreq_unregister_governor(&cpufreq_gov_ondemand);
 768 }
 769
 770
 771 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 772 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 773 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 774         "Low Latency Frequency Transition capable processors");
 775 MODULE_LICENSE("GPL");
 776
 777 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 778 fs_initcall(cpufreq_gov_dbs_init);
 779 #else
 780 module_init(cpufreq_gov_dbs_init);
 781 #endif
 782 module_exit(cpufreq_gov_dbs_exit);