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
  22 /*
  23  * dbs is used in this file as a shortform for demandbased switching
  24  * It helps to keep variable names smaller, simpler
  25  */
  26
  27 #define DEF_FREQUENCY_UP_THRESHOLD              (80)
  28 #define MIN_FREQUENCY_UP_THRESHOLD              (11)
  29 #define MAX_FREQUENCY_UP_THRESHOLD              (100)
  30
  31 /*
  32  * The polling frequency of this governor depends on the capability of
  33  * the processor. Default polling frequency is 1000 times the transition
  34  * latency of the processor. The governor will work on any processor with
  35  * transition latency <= 10mS, using appropriate sampling
  36  * rate.
  37  * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
  38  * this governor will not work.
  39  * All times here are in uS.
  40  */
  41 static unsigned int def_sampling_rate;
  42 #define MIN_SAMPLING_RATE_RATIO                 (2)
  43 /* for correct statistics, we need at least 10 ticks between each measure */
  44 #define MIN_STAT_SAMPLING_RATE                  (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
  45 #define MIN_SAMPLING_RATE                       (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
  46 #define MAX_SAMPLING_RATE                       (500 * def_sampling_rate)
  47 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER    (1000)
  48 #define TRANSITION_LATENCY_LIMIT                (10 * 1000)
  49
  50 static void do_dbs_timer(void *data);
  51
  52 struct cpu_dbs_info_s {
  53         cputime64_t prev_cpu_idle;
  54         cputime64_t prev_cpu_wall;
  55         struct cpufreq_policy *cur_policy;
  56         struct work_struct work;
  57         unsigned int enable;
  58         struct cpufreq_frequency_table *freq_table;
  59         unsigned int freq_lo;
  60         unsigned int freq_lo_jiffies;
  61         unsigned int freq_hi_jiffies;
  62 };
  63 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
  64
  65 static unsigned int dbs_enable; /* number of CPUs using this policy */
  66
  67 /*
  68  * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug
  69  * lock and dbs_mutex. cpu_hotplug lock should always be held before
  70  * dbs_mutex. If any function that can potentially take cpu_hotplug lock
  71  * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
  72  * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
  73  * is recursive for the same process. -Venki
  74  */
  75 static DEFINE_MUTEX(dbs_mutex);
  76
  77 static struct workqueue_struct  *kondemand_wq;
  78
  79 static struct dbs_tuners {
  80         unsigned int sampling_rate;
  81         unsigned int up_threshold;
  82         unsigned int ignore_nice;
  83         unsigned int powersave_bias;
  84 } dbs_tuners_ins = {
  85         .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
  86         .ignore_nice = 0,
  87         .powersave_bias = 0,
  88 };
  89
  90 static inline cputime64_t get_cpu_idle_time(unsigned int cpu)
  91 {
  92         cputime64_t retval;
  93
  94         retval = cputime64_add(kstat_cpu(cpu).cpustat.idle,
  95                         kstat_cpu(cpu).cpustat.iowait);
  96
  97         if (dbs_tuners_ins.ignore_nice)
  98                 retval = cputime64_add(retval, kstat_cpu(cpu).cpustat.nice);
  99
 100         return retval;
 101 }
 102
 103 /*
 104  * Find right freq to be set now with powersave_bias on.
 105  * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
 106  * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
 107  */
 108 unsigned int powersave_bias_target(struct cpufreq_policy *policy,
 109                 unsigned int freq_next, unsigned int relation)
 110 {
 111         unsigned int freq_req, freq_reduc, freq_avg;
 112         unsigned int freq_hi, freq_lo;
 113         unsigned int index = 0;
 114         unsigned int jiffies_total, jiffies_hi, jiffies_lo;
 115         struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, policy->cpu);
 116
 117         if (!dbs_info->freq_table) {
 118                 dbs_info->freq_lo = 0;
 119                 dbs_info->freq_lo_jiffies = 0;
 120                 return freq_next;
 121         }
 122
 123         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
 124                         relation, &index);
 125         freq_req = dbs_info->freq_table[index].frequency;
 126         freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
 127         freq_avg = freq_req - freq_reduc;
 128
 129         /* Find freq bounds for freq_avg in freq_table */
 130         index = 0;
 131         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 132                         CPUFREQ_RELATION_H, &index);
 133         freq_lo = dbs_info->freq_table[index].frequency;
 134         index = 0;
 135         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 136                         CPUFREQ_RELATION_L, &index);
 137         freq_hi = dbs_info->freq_table[index].frequency;
 138
 139         /* Find out how long we have to be in hi and lo freqs */
 140         if (freq_hi == freq_lo) {
 141                 dbs_info->freq_lo = 0;
 142                 dbs_info->freq_lo_jiffies = 0;
 143                 return freq_lo;
 144         }
 145         jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 146         jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
 147         jiffies_hi += ((freq_hi - freq_lo) / 2);
 148         jiffies_hi /= (freq_hi - freq_lo);
 149         jiffies_lo = jiffies_total - jiffies_hi;
 150         dbs_info->freq_lo = freq_lo;
 151         dbs_info->freq_lo_jiffies = jiffies_lo;
 152         dbs_info->freq_hi_jiffies = jiffies_hi;
 153         return freq_hi;
 154 }
 155
 156 static void ondemand_powersave_bias_init(void)
 157 {
 158         int i;
 159         for_each_online_cpu(i) {
 160                 struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, i);
 161                 dbs_info->freq_table = cpufreq_frequency_get_table(i);
 162                 dbs_info->freq_lo = 0;
 163         }
 164 }
 165
 166 /************************** sysfs interface ************************/
 167 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
 168 {
 169         return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
 170 }
 171
 172 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
 173 {
 174         return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
 175 }
 176
 177 #define define_one_ro(_name)            \
 178 static struct freq_attr _name =         \
 179 __ATTR(_name, 0444, show_##_name, NULL)
 180
 181 define_one_ro(sampling_rate_max);
 182 define_one_ro(sampling_rate_min);
 183
 184 /* cpufreq_ondemand Governor Tunables */
 185 #define show_one(file_name, object)                                     \
 186 static ssize_t show_##file_name                                         \
 187 (struct cpufreq_policy *unused, char *buf)                              \
 188 {                                                                       \
 189         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
 190 }
 191 show_one(sampling_rate, sampling_rate);
 192 show_one(up_threshold, up_threshold);
 193 show_one(ignore_nice_load, ignore_nice);
 194 show_one(powersave_bias, powersave_bias);
 195
 196 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
 197                 const char *buf, size_t count)
 198 {
 199         unsigned int input;
 200         int ret;
 201         ret = sscanf(buf, "%u", &input);
 202
 203         mutex_lock(&dbs_mutex);
 204         if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
 205                 mutex_unlock(&dbs_mutex);
 206                 return -EINVAL;
 207         }
 208
 209         dbs_tuners_ins.sampling_rate = input;
 210         mutex_unlock(&dbs_mutex);
 211
 212         return count;
 213 }
 214
 215 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
 216                 const char *buf, size_t count)
 217 {
 218         unsigned int input;
 219         int ret;
 220         ret = sscanf(buf, "%u", &input);
 221
 222         mutex_lock(&dbs_mutex);
 223         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 224                         input < MIN_FREQUENCY_UP_THRESHOLD) {
 225                 mutex_unlock(&dbs_mutex);
 226                 return -EINVAL;
 227         }
 228
 229         dbs_tuners_ins.up_threshold = input;
 230         mutex_unlock(&dbs_mutex);
 231
 232         return count;
 233 }
 234
 235 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
 236                 const char *buf, size_t count)
 237 {
 238         unsigned int input;
 239         int ret;
 240
 241         unsigned int j;
 242
 243         ret = sscanf(buf, "%u", &input);
 244         if ( ret != 1 )
 245                 return -EINVAL;
 246
 247         if ( input > 1 )
 248                 input = 1;
 249
 250         mutex_lock(&dbs_mutex);
 251         if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
 252                 mutex_unlock(&dbs_mutex);
 253                 return count;
 254         }
 255         dbs_tuners_ins.ignore_nice = input;
 256
 257         /* we need to re-evaluate prev_cpu_idle */
 258         for_each_online_cpu(j) {
 259                 struct cpu_dbs_info_s *dbs_info;
 260                 dbs_info = &per_cpu(cpu_dbs_info, j);
 261                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j);
 262                 dbs_info->prev_cpu_wall = get_jiffies_64();
 263         }
 264         mutex_unlock(&dbs_mutex);
 265
 266         return count;
 267 }
 268
 269 static ssize_t store_powersave_bias(struct cpufreq_policy *unused,
 270                 const char *buf, size_t count)
 271 {
 272         unsigned int input;
 273         int ret;
 274         ret = sscanf(buf, "%u", &input);
 275
 276         if (ret != 1)
 277                 return -EINVAL;
 278
 279         if (input > 1000)
 280                 input = 1000;
 281
 282         mutex_lock(&dbs_mutex);
 283         dbs_tuners_ins.powersave_bias = input;
 284         ondemand_powersave_bias_init();
 285         mutex_unlock(&dbs_mutex);
 286
 287         return count;
 288 }
 289
 290 #define define_one_rw(_name) \
 291 static struct freq_attr _name = \
 292 __ATTR(_name, 0644, show_##_name, store_##_name)
 293
 294 define_one_rw(sampling_rate);
 295 define_one_rw(up_threshold);
 296 define_one_rw(ignore_nice_load);
 297 define_one_rw(powersave_bias);
 298
 299 static struct attribute * dbs_attributes[] = {
 300         &sampling_rate_max.attr,
 301         &sampling_rate_min.attr,
 302         &sampling_rate.attr,
 303         &up_threshold.attr,
 304         &ignore_nice_load.attr,
 305         &powersave_bias.attr,
 306         NULL
 307 };
 308
 309 static struct attribute_group dbs_attr_group = {
 310         .attrs = dbs_attributes,
 311         .name = "ondemand",
 312 };
 313
 314 /************************** sysfs end ************************/
 315
 316 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
 317 {
 318         unsigned int idle_ticks, total_ticks;
 319         unsigned int load;
 320         cputime64_t cur_jiffies;
 321
 322         struct cpufreq_policy *policy;
 323         unsigned int j;
 324
 325         if (!this_dbs_info->enable)
 326                 return;
 327
 328         this_dbs_info->freq_lo = 0;
 329         policy = this_dbs_info->cur_policy;
 330         cur_jiffies = jiffies64_to_cputime64(get_jiffies_64());
 331         total_ticks = (unsigned int) cputime64_sub(cur_jiffies,
 332                         this_dbs_info->prev_cpu_wall);
 333         this_dbs_info->prev_cpu_wall = cur_jiffies;
 334         if (!total_ticks)
 335                 return;
 336         /*
 337          * Every sampling_rate, we check, if current idle time is less
 338          * than 20% (default), then we try to increase frequency
 339          * Every sampling_rate, we look for a the lowest
 340          * frequency which can sustain the load while keeping idle time over
 341          * 30%. If such a frequency exist, we try to decrease to this frequency.
 342          *
 343          * Any frequency increase takes it to the maximum frequency.
 344          * Frequency reduction happens at minimum steps of
 345          * 5% (default) of current frequency
 346          */
 347
 348         /* Get Idle Time */
 349         idle_ticks = UINT_MAX;
 350         for_each_cpu_mask(j, policy->cpus) {
 351                 cputime64_t total_idle_ticks;
 352                 unsigned int tmp_idle_ticks;
 353                 struct cpu_dbs_info_s *j_dbs_info;
 354
 355                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
 356                 total_idle_ticks = get_cpu_idle_time(j);
 357                 tmp_idle_ticks = (unsigned int) cputime64_sub(total_idle_ticks,
 358                                 j_dbs_info->prev_cpu_idle);
 359                 j_dbs_info->prev_cpu_idle = total_idle_ticks;
 360
 361                 if (tmp_idle_ticks < idle_ticks)
 362                         idle_ticks = tmp_idle_ticks;
 363         }
 364         load = (100 * (total_ticks - idle_ticks)) / total_ticks;
 365
 366         /* Check for frequency increase */
 367         if (load > dbs_tuners_ins.up_threshold) {
 368                 /* if we are already at full speed then break out early */
 369                 if (!dbs_tuners_ins.powersave_bias) {
 370                         if (policy->cur == policy->max)
 371                                 return;
 372
 373                         __cpufreq_driver_target(policy, policy->max,
 374                                 CPUFREQ_RELATION_H);
 375                 } else {
 376                         int freq = powersave_bias_target(policy, policy->max,
 377                                         CPUFREQ_RELATION_H);
 378                         __cpufreq_driver_target(policy, freq,
 379                                 CPUFREQ_RELATION_L);
 380                 }
 381                 return;
 382         }
 383
 384         /* Check for frequency decrease */
 385         /* if we cannot reduce the frequency anymore, break out early */
 386         if (policy->cur == policy->min)
 387                 return;
 388
 389         /*
 390          * The optimal frequency is the frequency that is the lowest that
 391          * can support the current CPU usage without triggering the up
 392          * policy. To be safe, we focus 10 points under the threshold.
 393          */
 394         if (load < (dbs_tuners_ins.up_threshold - 10)) {
 395                 unsigned int freq_next = (policy->cur * load) /
 396                         (dbs_tuners_ins.up_threshold - 10);
 397                 if (!dbs_tuners_ins.powersave_bias) {
 398                         __cpufreq_driver_target(policy, freq_next,
 399                                         CPUFREQ_RELATION_L);
 400                 } else {
 401                         int freq = powersave_bias_target(policy, freq_next,
 402                                         CPUFREQ_RELATION_L);
 403                         __cpufreq_driver_target(policy, freq,
 404                                 CPUFREQ_RELATION_L);
 405                 }
 406         }
 407 }
 408
 409 /* Sampling types */
 410 enum {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE};
 411
 412 static void do_dbs_timer(void *data)
 413 {
 414         unsigned int cpu = smp_processor_id();
 415         struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, cpu);
 416         /* We want all CPUs to do sampling nearly on same jiffy */
 417         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 418         delay -= jiffies % delay;
 419
 420         if (!dbs_info->enable)
 421                 return;
 422         /* Common NORMAL_SAMPLE setup */
 423         INIT_WORK(&dbs_info->work, do_dbs_timer, (void *)DBS_NORMAL_SAMPLE);
 424         if (!dbs_tuners_ins.powersave_bias ||
 425             (unsigned long) data == DBS_NORMAL_SAMPLE) {
 426                 lock_cpu_hotplug();
 427                 dbs_check_cpu(dbs_info);
 428                 unlock_cpu_hotplug();
 429                 if (dbs_info->freq_lo) {
 430                         /* Setup timer for SUB_SAMPLE */
 431                         INIT_WORK(&dbs_info->work, do_dbs_timer,
 432                                         (void *)DBS_SUB_SAMPLE);
 433                         delay = dbs_info->freq_hi_jiffies;
 434                 }
 435         } else {
 436                 __cpufreq_driver_target(dbs_info->cur_policy,
 437                                         dbs_info->freq_lo,
 438                                         CPUFREQ_RELATION_H);
 439         }
 440         queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
 441 }
 442
 443 static inline void dbs_timer_init(unsigned int cpu)
 444 {
 445         struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, cpu);
 446         /* We want all CPUs to do sampling nearly on same jiffy */
 447         int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 448         delay -= jiffies % delay;
 449
 450         ondemand_powersave_bias_init();
 451         INIT_WORK(&dbs_info->work, do_dbs_timer, 0);
 452         queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
 453 }
 454
 455 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
 456 {
 457         dbs_info->enable = 0;
 458         cancel_delayed_work(&dbs_info->work);
 459         flush_workqueue(kondemand_wq);
 460 }
 461
 462 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 463                                    unsigned int event)
 464 {
 465         unsigned int cpu = policy->cpu;
 466         struct cpu_dbs_info_s *this_dbs_info;
 467         unsigned int j;
 468
 469         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
 470
 471         switch (event) {
 472         case CPUFREQ_GOV_START:
 473                 if ((!cpu_online(cpu)) || (!policy->cur))
 474                         return -EINVAL;
 475
 476                 if (policy->cpuinfo.transition_latency >
 477                                 (TRANSITION_LATENCY_LIMIT * 1000)) {
 478                         printk(KERN_WARNING "ondemand governor failed to load "
 479                                "due to too long transition latency\n");
 480                         return -EINVAL;
 481                 }
 482                 if (this_dbs_info->enable) /* Already enabled */
 483                         break;
 484
 485                 mutex_lock(&dbs_mutex);
 486                 dbs_enable++;
 487                 if (dbs_enable == 1) {
 488                         kondemand_wq = create_workqueue("kondemand");
 489                         if (!kondemand_wq) {
 490                                 printk(KERN_ERR "Creation of kondemand failed\n");
 491                                 dbs_enable--;
 492                                 mutex_unlock(&dbs_mutex);
 493                                 return -ENOSPC;
 494                         }
 495                 }
 496                 for_each_cpu_mask(j, policy->cpus) {
 497                         struct cpu_dbs_info_s *j_dbs_info;
 498                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
 499                         j_dbs_info->cur_policy = policy;
 500
 501                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j);
 502                         j_dbs_info->prev_cpu_wall = get_jiffies_64();
 503                 }
 504                 this_dbs_info->enable = 1;
 505                 sysfs_create_group(&policy->kobj, &dbs_attr_group);
 506                 /*
 507                  * Start the timerschedule work, when this governor
 508                  * is used for first time
 509                  */
 510                 if (dbs_enable == 1) {
 511                         unsigned int latency;
 512                         /* policy latency is in nS. Convert it to uS first */
 513                         latency = policy->cpuinfo.transition_latency / 1000;
 514                         if (latency == 0)
 515                                 latency = 1;
 516
 517                         def_sampling_rate = latency *
 518                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
 519
 520                         if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
 521                                 def_sampling_rate = MIN_STAT_SAMPLING_RATE;
 522
 523                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
 524                 }
 525                 dbs_timer_init(policy->cpu);
 526
 527                 mutex_unlock(&dbs_mutex);
 528                 break;
 529
 530         case CPUFREQ_GOV_STOP:
 531                 mutex_lock(&dbs_mutex);
 532                 dbs_timer_exit(this_dbs_info);
 533                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
 534                 dbs_enable--;
 535                 if (dbs_enable == 0)
 536                         destroy_workqueue(kondemand_wq);
 537
 538                 mutex_unlock(&dbs_mutex);
 539
 540                 break;
 541
 542         case CPUFREQ_GOV_LIMITS:
 543                 mutex_lock(&dbs_mutex);
 544                 if (policy->max < this_dbs_info->cur_policy->cur)
 545                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 546                                                 policy->max,
 547                                                 CPUFREQ_RELATION_H);
 548                 else if (policy->min > this_dbs_info->cur_policy->cur)
 549                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 550                                                 policy->min,
 551                                                 CPUFREQ_RELATION_L);
 552                 mutex_unlock(&dbs_mutex);
 553                 break;
 554         }
 555         return 0;
 556 }
 557
 558 static struct cpufreq_governor cpufreq_gov_dbs = {
 559         .name = "ondemand",
 560         .governor = cpufreq_governor_dbs,
 561         .owner = THIS_MODULE,
 562 };
 563
 564 static int __init cpufreq_gov_dbs_init(void)
 565 {
 566         return cpufreq_register_governor(&cpufreq_gov_dbs);
 567 }
 568
 569 static void __exit cpufreq_gov_dbs_exit(void)
 570 {
 571         cpufreq_unregister_governor(&cpufreq_gov_dbs);
 572 }
 573
 574
 575 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 576 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 577 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 578                    "Low Latency Frequency Transition capable processors");
 579 MODULE_LICENSE("GPL");
 580
 581 module_init(cpufreq_gov_dbs_init);
 582 module_exit(cpufreq_gov_dbs_exit);