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 data in dbs_tuners_ins from concurrent changes on
 103  * different CPUs. It 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 cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
 124                                                         cputime64_t *wall)
 125 {
 126         cputime64_t idle_time;
 127         cputime64_t cur_wall_time;
 128         cputime64_t busy_time;
 129
 130         cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
 131         busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
 132                         kstat_cpu(cpu).cpustat.system);
 133
 134         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
 135         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
 136         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
 137         busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice);
 138
 139         idle_time = cputime64_sub(cur_wall_time, busy_time);
 140         if (wall)
 141                 *wall = (cputime64_t)jiffies_to_usecs(cur_wall_time);
 142
 143         return (cputime64_t)jiffies_to_usecs(idle_time);
 144 }
 145
 146 static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
 147 {
 148         u64 idle_time = get_cpu_idle_time_us(cpu, wall);
 149
 150         if (idle_time == -1ULL)
 151                 return get_cpu_idle_time_jiffy(cpu, wall);
 152
 153         return idle_time;
 154 }
 155
 156 static inline cputime64_t get_cpu_iowait_time(unsigned int cpu, cputime64_t *wall)
 157 {
 158         u64 iowait_time = get_cpu_iowait_time_us(cpu, wall);
 159
 160         if (iowait_time == -1ULL)
 161                 return 0;
 162
 163         return iowait_time;
 164 }
 165
 166 /*
 167  * Find right freq to be set now with powersave_bias on.
 168  * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
 169  * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
 170  */
 171 static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
 172                                           unsigned int freq_next,
 173                                           unsigned int relation)
 174 {
 175         unsigned int freq_req, freq_reduc, freq_avg;
 176         unsigned int freq_hi, freq_lo;
 177         unsigned int index = 0;
 178         unsigned int jiffies_total, jiffies_hi, jiffies_lo;
 179         struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
 180                                                    policy->cpu);
 181
 182         if (!dbs_info->freq_table) {
 183                 dbs_info->freq_lo = 0;
 184                 dbs_info->freq_lo_jiffies = 0;
 185                 return freq_next;
 186         }
 187
 188         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
 189                         relation, &index);
 190         freq_req = dbs_info->freq_table[index].frequency;
 191         freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
 192         freq_avg = freq_req - freq_reduc;
 193
 194         /* Find freq bounds for freq_avg in freq_table */
 195         index = 0;
 196         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 197                         CPUFREQ_RELATION_H, &index);
 198         freq_lo = dbs_info->freq_table[index].frequency;
 199         index = 0;
 200         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 201                         CPUFREQ_RELATION_L, &index);
 202         freq_hi = dbs_info->freq_table[index].frequency;
 203
 204         /* Find out how long we have to be in hi and lo freqs */
 205         if (freq_hi == freq_lo) {
 206                 dbs_info->freq_lo = 0;
 207                 dbs_info->freq_lo_jiffies = 0;
 208                 return freq_lo;
 209         }
 210         jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 211         jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
 212         jiffies_hi += ((freq_hi - freq_lo) / 2);
 213         jiffies_hi /= (freq_hi - freq_lo);
 214         jiffies_lo = jiffies_total - jiffies_hi;
 215         dbs_info->freq_lo = freq_lo;
 216         dbs_info->freq_lo_jiffies = jiffies_lo;
 217         dbs_info->freq_hi_jiffies = jiffies_hi;
 218         return freq_hi;
 219 }
 220
 221 static void ondemand_powersave_bias_init_cpu(int cpu)
 222 {
 223         struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 224         dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
 225         dbs_info->freq_lo = 0;
 226 }
 227
 228 static void ondemand_powersave_bias_init(void)
 229 {
 230         int i;
 231         for_each_online_cpu(i) {
 232                 ondemand_powersave_bias_init_cpu(i);
 233         }
 234 }
 235
 236 /************************** sysfs interface ************************/
 237
 238 static ssize_t show_sampling_rate_max(struct kobject *kobj,
 239                                       struct attribute *attr, char *buf)
 240 {
 241         printk_once(KERN_INFO "CPUFREQ: ondemand sampling_rate_max "
 242                "sysfs file is deprecated - used by: %s\n", current->comm);
 243         return sprintf(buf, "%u\n", -1U);
 244 }
 245
 246 static ssize_t show_sampling_rate_min(struct kobject *kobj,
 247                                       struct attribute *attr, char *buf)
 248 {
 249         return sprintf(buf, "%u\n", min_sampling_rate);
 250 }
 251
 252 define_one_global_ro(sampling_rate_max);
 253 define_one_global_ro(sampling_rate_min);
 254
 255 /* cpufreq_ondemand Governor Tunables */
 256 #define show_one(file_name, object)                                     \
 257 static ssize_t show_##file_name                                         \
 258 (struct kobject *kobj, struct attribute *attr, char *buf)              \
 259 {                                                                       \
 260         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
 261 }
 262 show_one(sampling_rate, sampling_rate);
 263 show_one(io_is_busy, io_is_busy);
 264 show_one(up_threshold, up_threshold);
 265 show_one(sampling_down_factor, sampling_down_factor);
 266 show_one(ignore_nice_load, ignore_nice);
 267 show_one(powersave_bias, powersave_bias);
 268
 269 /*** delete after deprecation time ***/
 270
 271 #define DEPRECATION_MSG(file_name)                                      \
 272         printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "       \
 273                     "interface is deprecated - " #file_name "\n");
 274
 275 #define show_one_old(file_name)                                         \
 276 static ssize_t show_##file_name##_old                                   \
 277 (struct cpufreq_policy *unused, char *buf)                              \
 278 {                                                                       \
 279         printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "       \
 280                     "interface is deprecated - " #file_name "\n");      \
 281         return show_##file_name(NULL, NULL, buf);                       \
 282 }
 283 show_one_old(sampling_rate);
 284 show_one_old(up_threshold);
 285 show_one_old(ignore_nice_load);
 286 show_one_old(powersave_bias);
 287 show_one_old(sampling_rate_min);
 288 show_one_old(sampling_rate_max);
 289
 290 cpufreq_freq_attr_ro_old(sampling_rate_min);
 291 cpufreq_freq_attr_ro_old(sampling_rate_max);
 292
 293 /*** delete after deprecation time ***/
 294
 295 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
 296                                    const char *buf, size_t count)
 297 {
 298         unsigned int input;
 299         int ret;
 300         ret = sscanf(buf, "%u", &input);
 301         if (ret != 1)
 302                 return -EINVAL;
 303
 304         mutex_lock(&dbs_mutex);
 305         dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
 306         mutex_unlock(&dbs_mutex);
 307
 308         return count;
 309 }
 310
 311 static ssize_t store_io_is_busy(struct kobject *a, struct attribute *b,
 312                                    const char *buf, size_t count)
 313 {
 314         unsigned int input;
 315         int ret;
 316
 317         ret = sscanf(buf, "%u", &input);
 318         if (ret != 1)
 319                 return -EINVAL;
 320
 321         mutex_lock(&dbs_mutex);
 322         dbs_tuners_ins.io_is_busy = !!input;
 323         mutex_unlock(&dbs_mutex);
 324
 325         return count;
 326 }
 327
 328 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
 329                                   const char *buf, size_t count)
 330 {
 331         unsigned int input;
 332         int ret;
 333         ret = sscanf(buf, "%u", &input);
 334
 335         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 336                         input < MIN_FREQUENCY_UP_THRESHOLD) {
 337                 return -EINVAL;
 338         }
 339
 340         mutex_lock(&dbs_mutex);
 341         dbs_tuners_ins.up_threshold = input;
 342         mutex_unlock(&dbs_mutex);
 343
 344         return count;
 345 }
 346
 347 static ssize_t store_sampling_down_factor(struct kobject *a,
 348                         struct attribute *b, const char *buf, size_t count)
 349 {
 350         unsigned int input, j;
 351         int ret;
 352         ret = sscanf(buf, "%u", &input);
 353
 354         if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
 355                 return -EINVAL;
 356         mutex_lock(&dbs_mutex);
 357         dbs_tuners_ins.sampling_down_factor = input;
 358
 359         /* Reset down sampling multiplier in case it was active */
 360         for_each_online_cpu(j) {
 361                 struct cpu_dbs_info_s *dbs_info;
 362                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
 363                 dbs_info->rate_mult = 1;
 364         }
 365         mutex_unlock(&dbs_mutex);
 366
 367         return count;
 368 }
 369
 370 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
 371                                       const char *buf, size_t count)
 372 {
 373         unsigned int input;
 374         int ret;
 375
 376         unsigned int j;
 377
 378         ret = sscanf(buf, "%u", &input);
 379         if (ret != 1)
 380                 return -EINVAL;
 381
 382         if (input > 1)
 383                 input = 1;
 384
 385         mutex_lock(&dbs_mutex);
 386         if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
 387                 mutex_unlock(&dbs_mutex);
 388                 return count;
 389         }
 390         dbs_tuners_ins.ignore_nice = input;
 391
 392         /* we need to re-evaluate prev_cpu_idle */
 393         for_each_online_cpu(j) {
 394                 struct cpu_dbs_info_s *dbs_info;
 395                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
 396                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 397                                                 &dbs_info->prev_cpu_wall);
 398                 if (dbs_tuners_ins.ignore_nice)
 399                         dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 400
 401         }
 402         mutex_unlock(&dbs_mutex);
 403
 404         return count;
 405 }
 406
 407 static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b,
 408                                     const char *buf, size_t count)
 409 {
 410         unsigned int input;
 411         int ret;
 412         ret = sscanf(buf, "%u", &input);
 413
 414         if (ret != 1)
 415                 return -EINVAL;
 416
 417         if (input > 1000)
 418                 input = 1000;
 419
 420         mutex_lock(&dbs_mutex);
 421         dbs_tuners_ins.powersave_bias = input;
 422         ondemand_powersave_bias_init();
 423         mutex_unlock(&dbs_mutex);
 424
 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_max.attr,
 437         &sampling_rate_min.attr,
 438         &sampling_rate.attr,
 439         &up_threshold.attr,
 440         &sampling_down_factor.attr,
 441         &ignore_nice_load.attr,
 442         &powersave_bias.attr,
 443         &io_is_busy.attr,
 444         NULL
 445 };
 446
 447 static struct attribute_group dbs_attr_group = {
 448         .attrs = dbs_attributes,
 449         .name = "ondemand",
 450 };
 451
 452 /*** delete after deprecation time ***/
 453
 454 #define write_one_old(file_name)                                        \
 455 static ssize_t store_##file_name##_old                                  \
 456 (struct cpufreq_policy *unused, const char *buf, size_t count)          \
 457 {                                                                       \
 458        printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs "        \
 459                    "interface is deprecated - " #file_name "\n");       \
 460        return store_##file_name(NULL, NULL, buf, count);                \
 461 }
 462 write_one_old(sampling_rate);
 463 write_one_old(up_threshold);
 464 write_one_old(ignore_nice_load);
 465 write_one_old(powersave_bias);
 466
 467 cpufreq_freq_attr_rw_old(sampling_rate);
 468 cpufreq_freq_attr_rw_old(up_threshold);
 469 cpufreq_freq_attr_rw_old(ignore_nice_load);
 470 cpufreq_freq_attr_rw_old(powersave_bias);
 471
 472 static struct attribute *dbs_attributes_old[] = {
 473        &sampling_rate_max_old.attr,
 474        &sampling_rate_min_old.attr,
 475        &sampling_rate_old.attr,
 476        &up_threshold_old.attr,
 477        &ignore_nice_load_old.attr,
 478        &powersave_bias_old.attr,
 479        NULL
 480 };
 481
 482 static struct attribute_group dbs_attr_group_old = {
 483        .attrs = dbs_attributes_old,
 484        .name = "ondemand",
 485 };
 486
 487 /*** delete after deprecation time ***/
 488
 489 /************************** sysfs end ************************/
 490
 491 static void dbs_freq_increase(struct cpufreq_policy *p, unsigned int freq)
 492 {
 493         if (dbs_tuners_ins.powersave_bias)
 494                 freq = powersave_bias_target(p, freq, CPUFREQ_RELATION_H);
 495         else if (p->cur == p->max)
 496                 return;
 497
 498         __cpufreq_driver_target(p, freq, dbs_tuners_ins.powersave_bias ?
 499                         CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
 500 }
 501
 502 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
 503 {
 504         unsigned int max_load_freq;
 505
 506         struct cpufreq_policy *policy;
 507         unsigned int j;
 508
 509         this_dbs_info->freq_lo = 0;
 510         policy = this_dbs_info->cur_policy;
 511
 512         /*
 513          * Every sampling_rate, we check, if current idle time is less
 514          * than 20% (default), then we try to increase frequency
 515          * Every sampling_rate, we look for a the lowest
 516          * frequency which can sustain the load while keeping idle time over
 517          * 30%. If such a frequency exist, we try to decrease to this frequency.
 518          *
 519          * Any frequency increase takes it to the maximum frequency.
 520          * Frequency reduction happens at minimum steps of
 521          * 5% (default) of current frequency
 522          */
 523
 524         /* Get Absolute Load - in terms of freq */
 525         max_load_freq = 0;
 526
 527         for_each_cpu(j, policy->cpus) {
 528                 struct cpu_dbs_info_s *j_dbs_info;
 529                 cputime64_t cur_wall_time, cur_idle_time, cur_iowait_time;
 530                 unsigned int idle_time, wall_time, iowait_time;
 531                 unsigned int load, load_freq;
 532                 int freq_avg;
 533
 534                 j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
 535
 536                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
 537                 cur_iowait_time = get_cpu_iowait_time(j, &cur_wall_time);
 538
 539                 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
 540                                 j_dbs_info->prev_cpu_wall);
 541                 j_dbs_info->prev_cpu_wall = cur_wall_time;
 542
 543                 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
 544                                 j_dbs_info->prev_cpu_idle);
 545                 j_dbs_info->prev_cpu_idle = cur_idle_time;
 546
 547                 iowait_time = (unsigned int) cputime64_sub(cur_iowait_time,
 548                                 j_dbs_info->prev_cpu_iowait);
 549                 j_dbs_info->prev_cpu_iowait = cur_iowait_time;
 550
 551                 if (dbs_tuners_ins.ignore_nice) {
 552                         cputime64_t cur_nice;
 553                         unsigned long cur_nice_jiffies;
 554
 555                         cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
 556                                          j_dbs_info->prev_cpu_nice);
 557                         /*
 558                          * Assumption: nice time between sampling periods will
 559                          * be less than 2^32 jiffies for 32 bit sys
 560                          */
 561                         cur_nice_jiffies = (unsigned long)
 562                                         cputime64_to_jiffies64(cur_nice);
 563
 564                         j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
 565                         idle_time += jiffies_to_usecs(cur_nice_jiffies);
 566                 }
 567
 568                 /*
 569                  * For the purpose of ondemand, waiting for disk IO is an
 570                  * indication that you're performance critical, and not that
 571                  * the system is actually idle. So subtract the iowait time
 572                  * from the cpu idle time.
 573                  */
 574
 575                 if (dbs_tuners_ins.io_is_busy && idle_time >= iowait_time)
 576                         idle_time -= iowait_time;
 577
 578                 if (unlikely(!wall_time || wall_time < idle_time))
 579                         continue;
 580
 581                 load = 100 * (wall_time - idle_time) / wall_time;
 582
 583                 freq_avg = __cpufreq_driver_getavg(policy, j);
 584                 if (freq_avg <= 0)
 585                         freq_avg = policy->cur;
 586
 587                 load_freq = load * freq_avg;
 588                 if (load_freq > max_load_freq)
 589                         max_load_freq = load_freq;
 590         }
 591
 592         /* Check for frequency increase */
 593         if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
 594                 /* If switching to max speed, apply sampling_down_factor */
 595                 if (policy->cur < policy->max)
 596                         this_dbs_info->rate_mult =
 597                                 dbs_tuners_ins.sampling_down_factor;
 598                 dbs_freq_increase(policy, policy->max);
 599                 return;
 600         }
 601
 602         /* Check for frequency decrease */
 603         /* if we cannot reduce the frequency anymore, break out early */
 604         if (policy->cur == policy->min)
 605                 return;
 606
 607         /*
 608          * The optimal frequency is the frequency that is the lowest that
 609          * can support the current CPU usage without triggering the up
 610          * policy. To be safe, we focus 10 points under the threshold.
 611          */
 612         if (max_load_freq <
 613             (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
 614              policy->cur) {
 615                 unsigned int freq_next;
 616                 freq_next = max_load_freq /
 617                                 (dbs_tuners_ins.up_threshold -
 618                                  dbs_tuners_ins.down_differential);
 619
 620                 /* No longer fully busy, reset rate_mult */
 621                 this_dbs_info->rate_mult = 1;
 622
 623                 if (freq_next < policy->min)
 624                         freq_next = policy->min;
 625
 626                 if (!dbs_tuners_ins.powersave_bias) {
 627                         __cpufreq_driver_target(policy, freq_next,
 628                                         CPUFREQ_RELATION_L);
 629                 } else {
 630                         int freq = powersave_bias_target(policy, freq_next,
 631                                         CPUFREQ_RELATION_L);
 632                         __cpufreq_driver_target(policy, freq,
 633                                 CPUFREQ_RELATION_L);
 634                 }
 635         }
 636 }
 637
 638 static void do_dbs_timer(struct work_struct *work)
 639 {
 640         struct cpu_dbs_info_s *dbs_info =
 641                 container_of(work, struct cpu_dbs_info_s, work.work);
 642         unsigned int cpu = dbs_info->cpu;
 643         int sample_type = dbs_info->sample_type;
 644
 645         /* We want all CPUs to do sampling nearly on same jiffy */
 646         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate
 647                 * dbs_info->rate_mult);
 648
 649         if (num_online_cpus() > 1)
 650                 delay -= jiffies % delay;
 651
 652         mutex_lock(&dbs_info->timer_mutex);
 653
 654         /* Common NORMAL_SAMPLE setup */
 655         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 656         if (!dbs_tuners_ins.powersave_bias ||
 657             sample_type == DBS_NORMAL_SAMPLE) {
 658                 dbs_check_cpu(dbs_info);
 659                 if (dbs_info->freq_lo) {
 660                         /* Setup timer for SUB_SAMPLE */
 661                         dbs_info->sample_type = DBS_SUB_SAMPLE;
 662                         delay = dbs_info->freq_hi_jiffies;
 663                 }
 664         } else {
 665                 __cpufreq_driver_target(dbs_info->cur_policy,
 666                         dbs_info->freq_lo, CPUFREQ_RELATION_H);
 667         }
 668         schedule_delayed_work_on(cpu, &dbs_info->work, delay);
 669         mutex_unlock(&dbs_info->timer_mutex);
 670 }
 671
 672 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
 673 {
 674         /* We want all CPUs to do sampling nearly on same jiffy */
 675         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 676
 677         if (num_online_cpus() > 1)
 678                 delay -= jiffies % delay;
 679
 680         dbs_info->sample_type = DBS_NORMAL_SAMPLE;
 681         INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
 682         schedule_delayed_work_on(dbs_info->cpu, &dbs_info->work, delay);
 683 }
 684
 685 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
 686 {
 687         cancel_delayed_work_sync(&dbs_info->work);
 688 }
 689
 690 /*
 691  * Not all CPUs want IO time to be accounted as busy; this dependson how
 692  * efficient idling at a higher frequency/voltage is.
 693  * Pavel Machek says this is not so for various generations of AMD and old
 694  * Intel systems.
 695  * Mike Chan (androidlcom) calis this is also not true for ARM.
 696  * Because of this, whitelist specific known (series) of CPUs by default, and
 697  * leave all others up to the user.
 698  */
 699 static int should_io_be_busy(void)
 700 {
 701 #if defined(CONFIG_X86)
 702         /*
 703          * For Intel, Core 2 (model 15) andl later have an efficient idle.
 704          */
 705         if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
 706             boot_cpu_data.x86 == 6 &&
 707             boot_cpu_data.x86_model >= 15)
 708                 return 1;
 709 #endif
 710         return 0;
 711 }
 712
 713 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 714                                    unsigned int event)
 715 {
 716         unsigned int cpu = policy->cpu;
 717         struct cpu_dbs_info_s *this_dbs_info;
 718         unsigned int j;
 719         int rc;
 720
 721         this_dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 722
 723         switch (event) {
 724         case CPUFREQ_GOV_START:
 725                 if ((!cpu_online(cpu)) || (!policy->cur))
 726                         return -EINVAL;
 727
 728                 mutex_lock(&dbs_mutex);
 729
 730                 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group_old);
 731                 if (rc) {
 732                         mutex_unlock(&dbs_mutex);
 733                         return rc;
 734                 }
 735
 736                 dbs_enable++;
 737                 for_each_cpu(j, policy->cpus) {
 738                         struct cpu_dbs_info_s *j_dbs_info;
 739                         j_dbs_info = &per_cpu(od_cpu_dbs_info, j);
 740                         j_dbs_info->cur_policy = policy;
 741
 742                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
 743                                                 &j_dbs_info->prev_cpu_wall);
 744                         if (dbs_tuners_ins.ignore_nice) {
 745                                 j_dbs_info->prev_cpu_nice =
 746                                                 kstat_cpu(j).cpustat.nice;
 747                         }
 748                 }
 749                 this_dbs_info->cpu = cpu;
 750                 this_dbs_info->rate_mult = 1;
 751                 ondemand_powersave_bias_init_cpu(cpu);
 752                 /*
 753                  * Start the timerschedule work, when this governor
 754                  * is used for first time
 755                  */
 756                 if (dbs_enable == 1) {
 757                         unsigned int latency;
 758
 759                         rc = sysfs_create_group(cpufreq_global_kobject,
 760                                                 &dbs_attr_group);
 761                         if (rc) {
 762                                 mutex_unlock(&dbs_mutex);
 763                                 return rc;
 764                         }
 765
 766                         /* policy latency is in nS. Convert it to uS first */
 767                         latency = policy->cpuinfo.transition_latency / 1000;
 768                         if (latency == 0)
 769                                 latency = 1;
 770                         /* Bring kernel and HW constraints together */
 771                         min_sampling_rate = max(min_sampling_rate,
 772                                         MIN_LATENCY_MULTIPLIER * latency);
 773                         dbs_tuners_ins.sampling_rate =
 774                                 max(min_sampling_rate,
 775                                     latency * LATENCY_MULTIPLIER);
 776                         dbs_tuners_ins.io_is_busy = should_io_be_busy();
 777                 }
 778                 mutex_unlock(&dbs_mutex);
 779
 780                 mutex_init(&this_dbs_info->timer_mutex);
 781                 dbs_timer_init(this_dbs_info);
 782                 break;
 783
 784         case CPUFREQ_GOV_STOP:
 785                 dbs_timer_exit(this_dbs_info);
 786
 787                 mutex_lock(&dbs_mutex);
 788                 sysfs_remove_group(&policy->kobj, &dbs_attr_group_old);
 789                 mutex_destroy(&this_dbs_info->timer_mutex);
 790                 dbs_enable--;
 791                 mutex_unlock(&dbs_mutex);
 792                 if (!dbs_enable)
 793                         sysfs_remove_group(cpufreq_global_kobject,
 794                                            &dbs_attr_group);
 795
 796                 break;
 797
 798         case CPUFREQ_GOV_LIMITS:
 799                 mutex_lock(&this_dbs_info->timer_mutex);
 800                 if (policy->max < this_dbs_info->cur_policy->cur)
 801                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 802                                 policy->max, CPUFREQ_RELATION_H);
 803                 else if (policy->min > this_dbs_info->cur_policy->cur)
 804                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 805                                 policy->min, CPUFREQ_RELATION_L);
 806                 mutex_unlock(&this_dbs_info->timer_mutex);
 807                 break;
 808         }
 809         return 0;
 810 }
 811
 812 static int __init cpufreq_gov_dbs_init(void)
 813 {
 814         cputime64_t wall;
 815         u64 idle_time;
 816         int cpu = get_cpu();
 817
 818         idle_time = get_cpu_idle_time_us(cpu, &wall);
 819         put_cpu();
 820         if (idle_time != -1ULL) {
 821                 /* Idle micro accounting is supported. Use finer thresholds */
 822                 dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
 823                 dbs_tuners_ins.down_differential =
 824                                         MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
 825                 /*
 826                  * In no_hz/micro accounting case we set the minimum frequency
 827                  * not depending on HZ, but fixed (very low). The deferred
 828                  * timer might skip some samples if idle/sleeping as needed.
 829                 */
 830                 min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
 831         } else {
 832                 /* For correct statistics, we need 10 ticks for each measure */
 833                 min_sampling_rate =
 834                         MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
 835         }
 836
 837         return cpufreq_register_governor(&cpufreq_gov_ondemand);
 838 }
 839
 840 static void __exit cpufreq_gov_dbs_exit(void)
 841 {
 842         cpufreq_unregister_governor(&cpufreq_gov_ondemand);
 843 }
 844
 845
 846 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 847 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 848 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 849         "Low Latency Frequency Transition capable processors");
 850 MODULE_LICENSE("GPL");
 851
 852 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 853 fs_initcall(cpufreq_gov_dbs_init);
 854 #else
 855 module_init(cpufreq_gov_dbs_init);
 856 #endif
 857 module_exit(cpufreq_gov_dbs_exit);