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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  14
  15 #include <linux/cpufreq.h>
  16 #include <linux/init.h>
  17 #include <linux/kernel.h>
  18 #include <linux/kernel_stat.h>
  19 #include <linux/kobject.h>
  20 #include <linux/module.h>
  21 #include <linux/mutex.h>
  22 #include <linux/percpu-defs.h>
  23 #include <linux/sysfs.h>
  24 #include <linux/tick.h>
  25 #include <linux/types.h>
  26
  27 #include "cpufreq_governor.h"
  28
  29 /* On-demand governor macors */
  30 #define DEF_FREQUENCY_DOWN_DIFFERENTIAL         (10)
  31 #define DEF_FREQUENCY_UP_THRESHOLD              (80)
  32 #define DEF_SAMPLING_DOWN_FACTOR                (1)
  33 #define MAX_SAMPLING_DOWN_FACTOR                (100000)
  34 #define MICRO_FREQUENCY_DOWN_DIFFERENTIAL       (3)
  35 #define MICRO_FREQUENCY_UP_THRESHOLD            (95)
  36 #define MICRO_FREQUENCY_MIN_SAMPLE_RATE         (10000)
  37 #define MIN_FREQUENCY_UP_THRESHOLD              (11)
  38 #define MAX_FREQUENCY_UP_THRESHOLD              (100)
  39
  40 static struct dbs_data od_dbs_data;
  41 static DEFINE_PER_CPU(struct od_cpu_dbs_info_s, od_cpu_dbs_info);
  42
  43 static struct od_dbs_tuners od_tuners = {
  44         .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
  45         .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
  46         .down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL,
  47         .ignore_nice = 0,
  48         .powersave_bias = 0,
  49 };
  50
  51 static void ondemand_powersave_bias_init_cpu(int cpu)
  52 {
  53         struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
  54
  55         dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
  56         dbs_info->freq_lo = 0;
  57 }
  58
  59 /*
  60  * Not all CPUs want IO time to be accounted as busy; this depends on how
  61  * efficient idling at a higher frequency/voltage is.
  62  * Pavel Machek says this is not so for various generations of AMD and old
  63  * Intel systems.
  64  * Mike Chan (androidlcom) calis this is also not true for ARM.
  65  * Because of this, whitelist specific known (series) of CPUs by default, and
  66  * leave all others up to the user.
  67  */
  68 static int should_io_be_busy(void)
  69 {
  70 #if defined(CONFIG_X86)
  71         /*
  72          * For Intel, Core 2 (model 15) andl later have an efficient idle.
  73          */
  74         if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
  75                         boot_cpu_data.x86 == 6 &&
  76                         boot_cpu_data.x86_model >= 15)
  77                 return 1;
  78 #endif
  79         return 0;
  80 }
  81
  82 /*
  83  * Find right freq to be set now with powersave_bias on.
  84  * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
  85  * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
  86  */
  87 static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
  88                 unsigned int freq_next, unsigned int relation)
  89 {
  90         unsigned int freq_req, freq_reduc, freq_avg;
  91         unsigned int freq_hi, freq_lo;
  92         unsigned int index = 0;
  93         unsigned int jiffies_total, jiffies_hi, jiffies_lo;
  94         struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
  95                                                    policy->cpu);
  96
  97         if (!dbs_info->freq_table) {
  98                 dbs_info->freq_lo = 0;
  99                 dbs_info->freq_lo_jiffies = 0;
 100                 return freq_next;
 101         }
 102
 103         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
 104                         relation, &index);
 105         freq_req = dbs_info->freq_table[index].frequency;
 106         freq_reduc = freq_req * od_tuners.powersave_bias / 1000;
 107         freq_avg = freq_req - freq_reduc;
 108
 109         /* Find freq bounds for freq_avg in freq_table */
 110         index = 0;
 111         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 112                         CPUFREQ_RELATION_H, &index);
 113         freq_lo = dbs_info->freq_table[index].frequency;
 114         index = 0;
 115         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 116                         CPUFREQ_RELATION_L, &index);
 117         freq_hi = dbs_info->freq_table[index].frequency;
 118
 119         /* Find out how long we have to be in hi and lo freqs */
 120         if (freq_hi == freq_lo) {
 121                 dbs_info->freq_lo = 0;
 122                 dbs_info->freq_lo_jiffies = 0;
 123                 return freq_lo;
 124         }
 125         jiffies_total = usecs_to_jiffies(od_tuners.sampling_rate);
 126         jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
 127         jiffies_hi += ((freq_hi - freq_lo) / 2);
 128         jiffies_hi /= (freq_hi - freq_lo);
 129         jiffies_lo = jiffies_total - jiffies_hi;
 130         dbs_info->freq_lo = freq_lo;
 131         dbs_info->freq_lo_jiffies = jiffies_lo;
 132         dbs_info->freq_hi_jiffies = jiffies_hi;
 133         return freq_hi;
 134 }
 135
 136 static void ondemand_powersave_bias_init(void)
 137 {
 138         int i;
 139         for_each_online_cpu(i) {
 140                 ondemand_powersave_bias_init_cpu(i);
 141         }
 142 }
 143
 144 static void dbs_freq_increase(struct cpufreq_policy *p, unsigned int freq)
 145 {
 146         if (od_tuners.powersave_bias)
 147                 freq = powersave_bias_target(p, freq, CPUFREQ_RELATION_H);
 148         else if (p->cur == p->max)
 149                 return;
 150
 151         __cpufreq_driver_target(p, freq, od_tuners.powersave_bias ?
 152                         CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
 153 }
 154
 155 /*
 156  * Every sampling_rate, we check, if current idle time is less than 20%
 157  * (default), then we try to increase frequency Every sampling_rate, we look for
 158  * a the lowest frequency which can sustain the load while keeping idle time
 159  * over 30%. If such a frequency exist, we try to decrease to this frequency.
 160  *
 161  * Any frequency increase takes it to the maximum frequency. Frequency reduction
 162  * happens at minimum steps of 5% (default) of current frequency
 163  */
 164 static void od_check_cpu(int cpu, unsigned int load_freq)
 165 {
 166         struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 167         struct cpufreq_policy *policy = dbs_info->cdbs.cur_policy;
 168
 169         dbs_info->freq_lo = 0;
 170
 171         /* Check for frequency increase */
 172         if (load_freq > od_tuners.up_threshold * policy->cur) {
 173                 /* If switching to max speed, apply sampling_down_factor */
 174                 if (policy->cur < policy->max)
 175                         dbs_info->rate_mult =
 176                                 od_tuners.sampling_down_factor;
 177                 dbs_freq_increase(policy, policy->max);
 178                 return;
 179         }
 180
 181         /* Check for frequency decrease */
 182         /* if we cannot reduce the frequency anymore, break out early */
 183         if (policy->cur == policy->min)
 184                 return;
 185
 186         /*
 187          * The optimal frequency is the frequency that is the lowest that can
 188          * support the current CPU usage without triggering the up policy. To be
 189          * safe, we focus 10 points under the threshold.
 190          */
 191         if (load_freq < (od_tuners.up_threshold - od_tuners.down_differential) *
 192                         policy->cur) {
 193                 unsigned int freq_next;
 194                 freq_next = load_freq / (od_tuners.up_threshold -
 195                                 od_tuners.down_differential);
 196
 197                 /* No longer fully busy, reset rate_mult */
 198                 dbs_info->rate_mult = 1;
 199
 200                 if (freq_next < policy->min)
 201                         freq_next = policy->min;
 202
 203                 if (!od_tuners.powersave_bias) {
 204                         __cpufreq_driver_target(policy, freq_next,
 205                                         CPUFREQ_RELATION_L);
 206                 } else {
 207                         int freq = powersave_bias_target(policy, freq_next,
 208                                         CPUFREQ_RELATION_L);
 209                         __cpufreq_driver_target(policy, freq,
 210                                         CPUFREQ_RELATION_L);
 211                 }
 212         }
 213 }
 214
 215 static void od_dbs_timer(struct work_struct *work)
 216 {
 217         struct od_cpu_dbs_info_s *dbs_info =
 218                 container_of(work, struct od_cpu_dbs_info_s, cdbs.work.work);
 219         unsigned int cpu = dbs_info->cdbs.cpu;
 220         int delay, sample_type = dbs_info->sample_type;
 221
 222         mutex_lock(&dbs_info->cdbs.timer_mutex);
 223
 224         /* Common NORMAL_SAMPLE setup */
 225         dbs_info->sample_type = OD_NORMAL_SAMPLE;
 226         if (sample_type == OD_SUB_SAMPLE) {
 227                 delay = dbs_info->freq_lo_jiffies;
 228                 __cpufreq_driver_target(dbs_info->cdbs.cur_policy,
 229                         dbs_info->freq_lo, CPUFREQ_RELATION_H);
 230         } else {
 231                 dbs_check_cpu(&od_dbs_data, cpu);
 232                 if (dbs_info->freq_lo) {
 233                         /* Setup timer for SUB_SAMPLE */
 234                         dbs_info->sample_type = OD_SUB_SAMPLE;
 235                         delay = dbs_info->freq_hi_jiffies;
 236                 } else {
 237                         delay = delay_for_sampling_rate(dbs_info->rate_mult);
 238                 }
 239         }
 240
 241         schedule_delayed_work_on(cpu, &dbs_info->cdbs.work, delay);
 242         mutex_unlock(&dbs_info->cdbs.timer_mutex);
 243 }
 244
 245 /************************** sysfs interface ************************/
 246
 247 static ssize_t show_sampling_rate_min(struct kobject *kobj,
 248                                       struct attribute *attr, char *buf)
 249 {
 250         return sprintf(buf, "%u\n", od_dbs_data.min_sampling_rate);
 251 }
 252
 253 /**
 254  * update_sampling_rate - update sampling rate effective immediately if needed.
 255  * @new_rate: new sampling rate
 256  *
 257  * If new rate is smaller than the old, simply updaing
 258  * dbs_tuners_int.sampling_rate might not be appropriate. For example, if the
 259  * original sampling_rate was 1 second and the requested new sampling rate is 10
 260  * ms because the user needs immediate reaction from ondemand governor, but not
 261  * sure if higher frequency will be required or not, then, the governor may
 262  * change the sampling rate too late; up to 1 second later. Thus, if we are
 263  * reducing the sampling rate, we need to make the new value effective
 264  * immediately.
 265  */
 266 static void update_sampling_rate(unsigned int new_rate)
 267 {
 268         int cpu;
 269
 270         od_tuners.sampling_rate = new_rate = max(new_rate,
 271                         od_dbs_data.min_sampling_rate);
 272
 273         for_each_online_cpu(cpu) {
 274                 struct cpufreq_policy *policy;
 275                 struct od_cpu_dbs_info_s *dbs_info;
 276                 unsigned long next_sampling, appointed_at;
 277
 278                 policy = cpufreq_cpu_get(cpu);
 279                 if (!policy)
 280                         continue;
 281                 dbs_info = &per_cpu(od_cpu_dbs_info, policy->cpu);
 282                 cpufreq_cpu_put(policy);
 283
 284                 mutex_lock(&dbs_info->cdbs.timer_mutex);
 285
 286                 if (!delayed_work_pending(&dbs_info->cdbs.work)) {
 287                         mutex_unlock(&dbs_info->cdbs.timer_mutex);
 288                         continue;
 289                 }
 290
 291                 next_sampling = jiffies + usecs_to_jiffies(new_rate);
 292                 appointed_at = dbs_info->cdbs.work.timer.expires;
 293
 294                 if (time_before(next_sampling, appointed_at)) {
 295
 296                         mutex_unlock(&dbs_info->cdbs.timer_mutex);
 297                         cancel_delayed_work_sync(&dbs_info->cdbs.work);
 298                         mutex_lock(&dbs_info->cdbs.timer_mutex);
 299
 300                         schedule_delayed_work_on(dbs_info->cdbs.cpu,
 301                                         &dbs_info->cdbs.work,
 302                                         usecs_to_jiffies(new_rate));
 303
 304                 }
 305                 mutex_unlock(&dbs_info->cdbs.timer_mutex);
 306         }
 307 }
 308
 309 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
 310                                    const char *buf, size_t count)
 311 {
 312         unsigned int input;
 313         int ret;
 314         ret = sscanf(buf, "%u", &input);
 315         if (ret != 1)
 316                 return -EINVAL;
 317         update_sampling_rate(input);
 318         return count;
 319 }
 320
 321 static ssize_t store_io_is_busy(struct kobject *a, struct attribute *b,
 322                                    const char *buf, size_t count)
 323 {
 324         unsigned int input;
 325         int ret;
 326
 327         ret = sscanf(buf, "%u", &input);
 328         if (ret != 1)
 329                 return -EINVAL;
 330         od_tuners.io_is_busy = !!input;
 331         return count;
 332 }
 333
 334 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
 335                                   const char *buf, size_t count)
 336 {
 337         unsigned int input;
 338         int ret;
 339         ret = sscanf(buf, "%u", &input);
 340
 341         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 342                         input < MIN_FREQUENCY_UP_THRESHOLD) {
 343                 return -EINVAL;
 344         }
 345         od_tuners.up_threshold = input;
 346         return count;
 347 }
 348
 349 static ssize_t store_sampling_down_factor(struct kobject *a,
 350                         struct attribute *b, const char *buf, size_t count)
 351 {
 352         unsigned int input, j;
 353         int ret;
 354         ret = sscanf(buf, "%u", &input);
 355
 356         if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
 357                 return -EINVAL;
 358         od_tuners.sampling_down_factor = input;
 359
 360         /* Reset down sampling multiplier in case it was active */
 361         for_each_online_cpu(j) {
 362                 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
 363                                 j);
 364                 dbs_info->rate_mult = 1;
 365         }
 366         return count;
 367 }
 368
 369 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
 370                                       const char *buf, size_t count)
 371 {
 372         unsigned int input;
 373         int ret;
 374
 375         unsigned int j;
 376
 377         ret = sscanf(buf, "%u", &input);
 378         if (ret != 1)
 379                 return -EINVAL;
 380
 381         if (input > 1)
 382                 input = 1;
 383
 384         if (input == od_tuners.ignore_nice) { /* nothing to do */
 385                 return count;
 386         }
 387         od_tuners.ignore_nice = input;
 388
 389         /* we need to re-evaluate prev_cpu_idle */
 390         for_each_online_cpu(j) {
 391                 struct od_cpu_dbs_info_s *dbs_info;
 392                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
 393                 dbs_info->cdbs.prev_cpu_idle = get_cpu_idle_time(j,
 394                                                 &dbs_info->cdbs.prev_cpu_wall);
 395                 if (od_tuners.ignore_nice)
 396                         dbs_info->cdbs.prev_cpu_nice =
 397                                 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 398
 399         }
 400         return count;
 401 }
 402
 403 static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b,
 404                                     const char *buf, size_t count)
 405 {
 406         unsigned int input;
 407         int ret;
 408         ret = sscanf(buf, "%u", &input);
 409
 410         if (ret != 1)
 411                 return -EINVAL;
 412
 413         if (input > 1000)
 414                 input = 1000;
 415
 416         od_tuners.powersave_bias = input;
 417         ondemand_powersave_bias_init();
 418         return count;
 419 }
 420
 421 show_one(od, sampling_rate, sampling_rate);
 422 show_one(od, io_is_busy, io_is_busy);
 423 show_one(od, up_threshold, up_threshold);
 424 show_one(od, sampling_down_factor, sampling_down_factor);
 425 show_one(od, ignore_nice_load, ignore_nice);
 426 show_one(od, powersave_bias, powersave_bias);
 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 define_one_global_ro(sampling_rate_min);
 435
 436 static struct attribute *dbs_attributes[] = {
 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 od_attr_group = {
 448         .attrs = dbs_attributes,
 449         .name = "ondemand",
 450 };
 451
 452 /************************** sysfs end ************************/
 453
 454 define_get_cpu_dbs_routines(od_cpu_dbs_info);
 455
 456 static struct od_ops od_ops = {
 457         .io_busy = should_io_be_busy,
 458         .powersave_bias_init_cpu = ondemand_powersave_bias_init_cpu,
 459         .powersave_bias_target = powersave_bias_target,
 460         .freq_increase = dbs_freq_increase,
 461 };
 462
 463 static struct dbs_data od_dbs_data = {
 464         .governor = GOV_ONDEMAND,
 465         .attr_group = &od_attr_group,
 466         .tuners = &od_tuners,
 467         .get_cpu_cdbs = get_cpu_cdbs,
 468         .get_cpu_dbs_info_s = get_cpu_dbs_info_s,
 469         .gov_dbs_timer = od_dbs_timer,
 470         .gov_check_cpu = od_check_cpu,
 471         .gov_ops = &od_ops,
 472 };
 473
 474 static int od_cpufreq_governor_dbs(struct cpufreq_policy *policy,
 475                 unsigned int event)
 476 {
 477         return cpufreq_governor_dbs(&od_dbs_data, policy, event);
 478 }
 479
 480 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 481 static
 482 #endif
 483 struct cpufreq_governor cpufreq_gov_ondemand = {
 484         .name                   = "ondemand",
 485         .governor               = od_cpufreq_governor_dbs,
 486         .max_transition_latency = TRANSITION_LATENCY_LIMIT,
 487         .owner                  = THIS_MODULE,
 488 };
 489
 490 static int __init cpufreq_gov_dbs_init(void)
 491 {
 492         u64 idle_time;
 493         int cpu = get_cpu();
 494
 495         mutex_init(&od_dbs_data.mutex);
 496         idle_time = get_cpu_idle_time_us(cpu, NULL);
 497         put_cpu();
 498         if (idle_time != -1ULL) {
 499                 /* Idle micro accounting is supported. Use finer thresholds */
 500                 od_tuners.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
 501                 od_tuners.down_differential = MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
 502                 /*
 503                  * In nohz/micro accounting case we set the minimum frequency
 504                  * not depending on HZ, but fixed (very low). The deferred
 505                  * timer might skip some samples if idle/sleeping as needed.
 506                 */
 507                 od_dbs_data.min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
 508         } else {
 509                 /* For correct statistics, we need 10 ticks for each measure */
 510                 od_dbs_data.min_sampling_rate = MIN_SAMPLING_RATE_RATIO *
 511                         jiffies_to_usecs(10);
 512         }
 513
 514         return cpufreq_register_governor(&cpufreq_gov_ondemand);
 515 }
 516
 517 static void __exit cpufreq_gov_dbs_exit(void)
 518 {
 519         cpufreq_unregister_governor(&cpufreq_gov_ondemand);
 520 }
 521
 522 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 523 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 524 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 525         "Low Latency Frequency Transition capable processors");
 526 MODULE_LICENSE("GPL");
 527
 528 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 529 fs_initcall(cpufreq_gov_dbs_init);
 530 #else
 531 module_init(cpufreq_gov_dbs_init);
 532 #endif
 533 module_exit(cpufreq_gov_dbs_exit);