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 /**
 262  * update_sampling_rate - update sampling rate effective immediately if needed.
 263  * @new_rate: new sampling rate
 264  *
 265  * If new rate is smaller than the old, simply updaing
 266  * dbs_tuners_int.sampling_rate might not be appropriate. For example,
 267  * if the original sampling_rate was 1 second and the requested new sampling
 268  * rate is 10 ms because the user needs immediate reaction from ondemand
 269  * governor, but not sure if higher frequency will be required or not,
 270  * then, the governor may change the sampling rate too late; up to 1 second
 271  * later. Thus, if we are reducing the sampling rate, we need to make the
 272  * new value effective immediately.
 273  */
 274 static void update_sampling_rate(unsigned int new_rate)
 275 {
 276         int cpu;
 277
 278         dbs_tuners_ins.sampling_rate = new_rate
 279                                      = max(new_rate, min_sampling_rate);
 280
 281         for_each_online_cpu(cpu) {
 282                 struct cpufreq_policy *policy;
 283                 struct cpu_dbs_info_s *dbs_info;
 284                 unsigned long next_sampling, appointed_at;
 285
 286                 policy = cpufreq_cpu_get(cpu);
 287                 if (!policy)
 288                         continue;
 289                 dbs_info = &per_cpu(od_cpu_dbs_info, policy->cpu);
 290                 cpufreq_cpu_put(policy);
 291
 292                 mutex_lock(&dbs_info->timer_mutex);
 293
 294                 if (!delayed_work_pending(&dbs_info->work)) {
 295                         mutex_unlock(&dbs_info->timer_mutex);
 296                         continue;
 297                 }
 298
 299                 next_sampling  = jiffies + usecs_to_jiffies(new_rate);
 300                 appointed_at = dbs_info->work.timer.expires;
 301
 302
 303                 if (time_before(next_sampling, appointed_at)) {
 304
 305                         mutex_unlock(&dbs_info->timer_mutex);
 306                         cancel_delayed_work_sync(&dbs_info->work);
 307                         mutex_lock(&dbs_info->timer_mutex);
 308
 309                         schedule_delayed_work_on(dbs_info->cpu, &dbs_info->work,
 310                                                  usecs_to_jiffies(new_rate));
 311
 312                 }
 313                 mutex_unlock(&dbs_info->timer_mutex);
 314         }
 315 }
 316
 317 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
 318                                    const char *buf, size_t count)
 319 {
 320         unsigned int input;
 321         int ret;
 322         ret = sscanf(buf, "%u", &input);
 323         if (ret != 1)
 324                 return -EINVAL;
 325         update_sampling_rate(input);
 326         return count;
 327 }
 328
 329 static ssize_t store_io_is_busy(struct kobject *a, struct attribute *b,
 330                                    const char *buf, size_t count)
 331 {
 332         unsigned int input;
 333         int ret;
 334
 335         ret = sscanf(buf, "%u", &input);
 336         if (ret != 1)
 337                 return -EINVAL;
 338         dbs_tuners_ins.io_is_busy = !!input;
 339         return count;
 340 }
 341
 342 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
 343                                   const char *buf, size_t count)
 344 {
 345         unsigned int input;
 346         int ret;
 347         ret = sscanf(buf, "%u", &input);
 348
 349         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 350                         input < MIN_FREQUENCY_UP_THRESHOLD) {
 351                 return -EINVAL;
 352         }
 353         dbs_tuners_ins.up_threshold = input;
 354         return count;
 355 }
 356
 357 static ssize_t store_sampling_down_factor(struct kobject *a,
 358                         struct attribute *b, const char *buf, size_t count)
 359 {
 360         unsigned int input, j;
 361         int ret;
 362         ret = sscanf(buf, "%u", &input);
 363
 364         if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
 365                 return -EINVAL;
 366         dbs_tuners_ins.sampling_down_factor = input;
 367
 368         /* Reset down sampling multiplier in case it was active */
 369         for_each_online_cpu(j) {
 370                 struct cpu_dbs_info_s *dbs_info;
 371                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
 372                 dbs_info->rate_mult = 1;
 373         }
 374         return count;
 375 }
 376
 377 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
 378                                       const char *buf, size_t count)
 379 {
 380         unsigned int input;
 381         int ret;
 382
 383         unsigned int j;
 384
 385         ret = sscanf(buf, "%u", &input);
 386         if (ret != 1)
 387                 return -EINVAL;
 388
 389         if (input > 1)
 390                 input = 1;
 391
 392         if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
 393                 return count;
 394         }
 395         dbs_tuners_ins.ignore_nice = input;
 396
 397         /* we need to re-evaluate prev_cpu_idle */
 398         for_each_online_cpu(j) {
 399                 struct cpu_dbs_info_s *dbs_info;
 400                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
 401                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 402                                                 &dbs_info->prev_cpu_wall);
 403                 if (dbs_tuners_ins.ignore_nice)
 404                         dbs_info->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 405
 406         }
 407         return count;
 408 }
 409
 410 static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b,
 411                                     const char *buf, size_t count)
 412 {
 413         unsigned int input;
 414         int ret;
 415         ret = sscanf(buf, "%u", &input);
 416
 417         if (ret != 1)
 418                 return -EINVAL;
 419
 420         if (input > 1000)
 421                 input = 1000;
 422
 423         dbs_tuners_ins.powersave_bias = input;
 424         ondemand_powersave_bias_init();
 425         return count;
 426 }
 427
 428 define_one_global_rw(sampling_rate);
 429 define_one_global_rw(io_is_busy);
 430 define_one_global_rw(up_threshold);
 431 define_one_global_rw(sampling_down_factor);
 432 define_one_global_rw(ignore_nice_load);
 433 define_one_global_rw(powersave_bias);
 434
 435 static struct attribute *dbs_attributes[] = {
 436         &sampling_rate_min.attr,
 437         &sampling_rate.attr,
 438         &up_threshold.attr,
 439         &sampling_down_factor.attr,
 440         &ignore_nice_load.attr,
 441         &powersave_bias.attr,
 442         &io_is_busy.attr,
 443         NULL
 444 };
 445
 446 static struct attribute_group dbs_attr_group = {
 447         .attrs = dbs_attributes,
 448         .name = "ondemand",
 449 };
 450
 451 /************************** sysfs end ************************/
 452
 453 static void dbs_freq_increase(struct cpufreq_policy *p, unsigned int freq)
 454 {
 455         if (dbs_tuners_ins.powersave_bias)
 456                 freq = powersave_bias_target(p, freq, CPUFREQ_RELATION_H);
 457         else if (p->cur == p->max)
 458                 return;
 459
 460         __cpufreq_driver_target(p, freq, dbs_tuners_ins.powersave_bias ?
 461                         CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
 462 }
 463
 464 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
 465 {
 466         unsigned int max_load_freq;
 467
 468         struct cpufreq_policy *policy;
 469         unsigned int j;
 470
 471         this_dbs_info->freq_lo = 0;
 472         policy = this_dbs_info->cur_policy;
 473
 474         /*
 475          * Every sampling_rate, we check, if current idle time is less
 476          * than 20% (default), then we try to increase frequency
 477          * Every sampling_rate, we look for a the lowest
 478          * frequency which can sustain the load while keeping idle time over
 479          * 30%. If such a frequency exist, we try to decrease to this frequency.
 480          *
 481          * Any frequency increase takes it to the maximum frequency.
 482          * Frequency reduction happens at minimum steps of
 483          * 5% (default) of current frequency
 484          */
 485
 486         /* Get Absolute Load - in terms of freq */
 487         max_load_freq = 0;
 488
 489         for_each_cpu(j, policy->cpus) {
 490                 struct cpu_dbs_info_s *j_dbs_info;
 491                 cputime64_t cur_wall_time, cur_idle_time, cur_iowait_time;
 492                 unsigned int idle_time, wall_time, iowait_time;
 493                 unsigned int load, load_freq;
 494                 int freq_avg;
 495
 496                 j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
 497
 498                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
 499                 cur_iowait_time = get_cpu_iowait_time(j, &cur_wall_time);
 500
 501                 wall_time = (unsigned int)
 502                         (cur_wall_time - j_dbs_info->prev_cpu_wall);
 503                 j_dbs_info->prev_cpu_wall = cur_wall_time;
 504
 505                 idle_time = (unsigned int)
 506                         (cur_idle_time - j_dbs_info->prev_cpu_idle);
 507                 j_dbs_info->prev_cpu_idle = cur_idle_time;
 508
 509                 iowait_time = (unsigned int)
 510                         (cur_iowait_time - j_dbs_info->prev_cpu_iowait);
 511                 j_dbs_info->prev_cpu_iowait = cur_iowait_time;
 512
 513                 if (dbs_tuners_ins.ignore_nice) {
 514                         u64 cur_nice;
 515                         unsigned long cur_nice_jiffies;
 516
 517                         cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] -
 518                                          j_dbs_info->prev_cpu_nice;
 519                         /*
 520                          * Assumption: nice time between sampling periods will
 521                          * be less than 2^32 jiffies for 32 bit sys
 522                          */
 523                         cur_nice_jiffies = (unsigned long)
 524                                         cputime64_to_jiffies64(cur_nice);
 525
 526                         j_dbs_info->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 527                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
 528                 }
 529
 530                 /*
 531                  * For the purpose of ondemand, waiting for disk IO is an
 532                  * indication that you're performance critical, and not that
 533                  * the system is actually idle. So subtract the iowait time
 534                  * from the cpu idle time.
 535                  */
 536
 537                 if (dbs_tuners_ins.io_is_busy && idle_time >= iowait_time)
 538                         idle_time -= iowait_time;
 539
 540                 if (unlikely(!wall_time || wall_time < idle_time))
 541                         continue;
 542
 543                 load = 100 * (wall_time - idle_time) / wall_time;
 544
 545                 freq_avg = __cpufreq_driver_getavg(policy, j);
 546                 if (freq_avg <= 0)
 547                         freq_avg = policy->cur;
 548
 549                 load_freq = load * freq_avg;
 550                 if (load_freq > max_load_freq)
 551                         max_load_freq = load_freq;
 552         }
 553
 554         /* Check for frequency increase */
 555         if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
 556                 /* If switching to max speed, apply sampling_down_factor */
 557                 if (policy->cur < policy->max)
 558                         this_dbs_info->rate_mult =
 559                                 dbs_tuners_ins.sampling_down_factor;
 560                 dbs_freq_increase(policy, policy->max);
 561                 return;
 562         }
 563
 564         /* Check for frequency decrease */
 565         /* if we cannot reduce the frequency anymore, break out early */
 566         if (policy->cur == policy->min)
 567                 return;
 568
 569         /*
 570          * The optimal frequency is the frequency that is the lowest that
 571          * can support the current CPU usage without triggering the up
 572          * policy. To be safe, we focus 10 points under the threshold.
 573          */
 574         if (max_load_freq <
 575             (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
 576              policy->cur) {
 577                 unsigned int freq_next;
 578                 freq_next = max_load_freq /
 579                                 (dbs_tuners_ins.up_threshold -
 580                                  dbs_tuners_ins.down_differential);
 581
 582                 /* No longer fully busy, reset rate_mult */
 583                 this_dbs_info->rate_mult = 1;
 584
 585                 if (freq_next < policy->min)
 586                         freq_next = policy->min;
 587
 588                 if (!dbs_tuners_ins.powersave_bias) {
 589                         __cpufreq_driver_target(policy, freq_next,
 590                                         CPUFREQ_RELATION_L);
 591                 } else {
 592                         int freq = powersave_bias_target(policy, freq_next,
 593                                         CPUFREQ_RELATION_L);
 594                         __cpufreq_driver_target(policy, freq,
 595                                 CPUFREQ_RELATION_L);
 596                 }
 597         }
 598 }
 599
 600 static void do_dbs_timer(struct work_struct *work)
 601 {
 602         struct cpu_dbs_info_s *dbs_info =
 603                 container_of(work, struct cpu_dbs_info_s, work.work);
 604         unsigned int cpu = dbs_info->cpu;
 605         int sample_type = dbs_info->sample_type;
 606
 607         int delay;
 608
 609         mutex_lock(&dbs_info->timer_mutex);
 610
 611         /* Common NORMAL_SAMPLE setup */
 612         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 613         if (!dbs_tuners_ins.powersave_bias ||
 614             sample_type == DBS_NORMAL_SAMPLE) {
 615                 dbs_check_cpu(dbs_info);
 616                 if (dbs_info->freq_lo) {
 617                         /* Setup timer for SUB_SAMPLE */
 618                         dbs_info->sample_type = DBS_SUB_SAMPLE;
 619                         delay = dbs_info->freq_hi_jiffies;
 620                 } else {
 621                         /* We want all CPUs to do sampling nearly on
 622                          * same jiffy
 623                          */
 624                         delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate
 625                                 * dbs_info->rate_mult);
 626
 627                         if (num_online_cpus() > 1)
 628                                 delay -= jiffies % delay;
 629                 }
 630         } else {
 631                 __cpufreq_driver_target(dbs_info->cur_policy,
 632                         dbs_info->freq_lo, CPUFREQ_RELATION_H);
 633                 delay = dbs_info->freq_lo_jiffies;
 634         }
 635         schedule_delayed_work_on(cpu, &dbs_info->work, delay);
 636         mutex_unlock(&dbs_info->timer_mutex);
 637 }
 638
 639 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
 640 {
 641         /* We want all CPUs to do sampling nearly on same jiffy */
 642         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 643
 644         if (num_online_cpus() > 1)
 645                 delay -= jiffies % delay;
 646
 647         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 648         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
 649         schedule_delayed_work_on(dbs_info->cpu, &dbs_info->work, delay);
 650 }
 651
 652 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
 653 {
 654         cancel_delayed_work_sync(&dbs_info->work);
 655 }
 656
 657 /*
 658  * Not all CPUs want IO time to be accounted as busy; this dependson how
 659  * efficient idling at a higher frequency/voltage is.
 660  * Pavel Machek says this is not so for various generations of AMD and old
 661  * Intel systems.
 662  * Mike Chan (androidlcom) calis this is also not true for ARM.
 663  * Because of this, whitelist specific known (series) of CPUs by default, and
 664  * leave all others up to the user.
 665  */
 666 static int should_io_be_busy(void)
 667 {
 668 #if defined(CONFIG_X86)
 669         /*
 670          * For Intel, Core 2 (model 15) andl later have an efficient idle.
 671          */
 672         if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
 673             boot_cpu_data.x86 == 6 &&
 674             boot_cpu_data.x86_model >= 15)
 675                 return 1;
 676 #endif
 677         return 0;
 678 }
 679
 680 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 681                                    unsigned int event)
 682 {
 683         unsigned int cpu = policy->cpu;
 684         struct cpu_dbs_info_s *this_dbs_info;
 685         unsigned int j;
 686         int rc;
 687
 688         this_dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 689
 690         switch (event) {
 691         case CPUFREQ_GOV_START:
 692                 if ((!cpu_online(cpu)) || (!policy->cur))
 693                         return -EINVAL;
 694
 695                 mutex_lock(&dbs_mutex);
 696
 697                 dbs_enable++;
 698                 for_each_cpu(j, policy->cpus) {
 699                         struct cpu_dbs_info_s *j_dbs_info;
 700                         j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
 701                         j_dbs_info->cur_policy = policy;
 702
 703                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 704                                                 &j_dbs_info->prev_cpu_wall);
 705                         if (dbs_tuners_ins.ignore_nice)
 706                                 j_dbs_info->prev_cpu_nice =
 707                                                 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 708                 }
 709                 this_dbs_info->cpu = cpu;
 710                 this_dbs_info->rate_mult = 1;
 711                 ondemand_powersave_bias_init_cpu(cpu);
 712                 /*
 713                  * Start the timerschedule work, when this governor
 714                  * is used for first time
 715                  */
 716                 if (dbs_enable == 1) {
 717                         unsigned int latency;
 718
 719                         rc = sysfs_create_group(cpufreq_global_kobject,
 720                                                 &dbs_attr_group);
 721                         if (rc) {
 722                                 mutex_unlock(&dbs_mutex);
 723                                 return rc;
 724                         }
 725
 726                         /* policy latency is in nS. Convert it to uS first */
 727                         latency = policy->cpuinfo.transition_latency / 1000;
 728                         if (latency == 0)
 729                                 latency = 1;
 730                         /* Bring kernel and HW constraints together */
 731                         min_sampling_rate = max(min_sampling_rate,
 732                                         MIN_LATENCY_MULTIPLIER * latency);
 733                         dbs_tuners_ins.sampling_rate =
 734                                 max(min_sampling_rate,
 735                                     latency * LATENCY_MULTIPLIER);
 736                         /*
 737                          * Cut def_sampling rate to 300ms if it was above,
 738                          * still consider to not set it above latency
 739                          * transition * 100
 740                          */
 741                         if (dbs_tuners_ins.sampling_rate > MAX_DEFAULT_SAMPLING_RATE) {
 742                                 dbs_tuners_ins.sampling_rate =
 743                                         max(min_sampling_rate, MAX_DEFAULT_SAMPLING_RATE);
 744                                 printk(KERN_INFO "CPUFREQ: ondemand sampling "
 745                                        "rate set to %d ms\n",
 746                                        dbs_tuners_ins.sampling_rate / 1000);
 747                         }
 748                         /*
 749                          * Be conservative in respect to performance.
 750                          * If an application calculates using two threads
 751                          * depending on each other, they will be run on several
 752                          * CPU cores resulting on 50% load on both.
 753                          * SLED might still want to prefer 80% up_threshold
 754                          * by default, but we cannot differ that here.
 755                          */
 756                         if (num_online_cpus() > 1)
 757                                 dbs_tuners_ins.up_threshold =
 758                                         DEF_FREQUENCY_UP_THRESHOLD / 2;
 759                         dbs_tuners_ins.io_is_busy = should_io_be_busy();
 760                 }
 761                 mutex_unlock(&dbs_mutex);
 762
 763                 mutex_init(&this_dbs_info->timer_mutex);
 764                 dbs_timer_init(this_dbs_info);
 765                 break;
 766
 767         case CPUFREQ_GOV_STOP:
 768                 dbs_timer_exit(this_dbs_info);
 769
 770                 mutex_lock(&dbs_mutex);
 771                 mutex_destroy(&this_dbs_info->timer_mutex);
 772                 dbs_enable--;
 773                 mutex_unlock(&dbs_mutex);
 774                 if (!dbs_enable)
 775                         sysfs_remove_group(cpufreq_global_kobject,
 776                                            &dbs_attr_group);
 777
 778                 break;
 779
 780         case CPUFREQ_GOV_LIMITS:
 781                 mutex_lock(&this_dbs_info->timer_mutex);
 782                 if (policy->max < this_dbs_info->cur_policy->cur)
 783                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 784                                 policy->max, CPUFREQ_RELATION_H);
 785                 else if (policy->min > this_dbs_info->cur_policy->cur)
 786                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 787                                 policy->min, CPUFREQ_RELATION_L);
 788                 mutex_unlock(&this_dbs_info->timer_mutex);
 789                 break;
 790         }
 791         return 0;
 792 }
 793
 794 static int __init cpufreq_gov_dbs_init(void)
 795 {
 796         u64 idle_time;
 797         int cpu = get_cpu();
 798
 799         idle_time = get_cpu_idle_time_us(cpu, NULL);
 800         put_cpu();
 801         if (idle_time != -1ULL) {
 802                 /* Idle micro accounting is supported. Use finer thresholds */
 803                 dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
 804                 dbs_tuners_ins.down_differential =
 805                                         MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
 806                 /*
 807                  * In nohz/micro accounting case we set the minimum frequency
 808                  * not depending on HZ, but fixed (very low). The deferred
 809                  * timer might skip some samples if idle/sleeping as needed.
 810                 */
 811                 min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
 812         } else {
 813                 /* For correct statistics, we need 10 ticks for each measure */
 814                 min_sampling_rate =
 815                         MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
 816         }
 817
 818         return cpufreq_register_governor(&cpufreq_gov_ondemand);
 819 }
 820
 821 static void __exit cpufreq_gov_dbs_exit(void)
 822 {
 823         cpufreq_unregister_governor(&cpufreq_gov_ondemand);
 824 }
 825
 826
 827 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 828 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 829 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 830         "Low Latency Frequency Transition capable processors");
 831 MODULE_LICENSE("GPL");
 832
 833 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 834 fs_initcall(cpufreq_gov_dbs_init);
 835 #else
 836 module_init(cpufreq_gov_dbs_init);
 837 #endif
 838 module_exit(cpufreq_gov_dbs_exit);