linux/drivers/cpufreq/cpufreq_ondemand.c
<<
>>
Prefs
   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 macros */
  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
  40static struct dbs_data od_dbs_data;
  41static DEFINE_PER_CPU(struct od_cpu_dbs_info_s, od_cpu_dbs_info);
  42
  43#ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
  44static struct cpufreq_governor cpufreq_gov_ondemand;
  45#endif
  46
  47static struct od_dbs_tuners od_tuners = {
  48        .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
  49        .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR,
  50        .adj_up_threshold = DEF_FREQUENCY_UP_THRESHOLD -
  51                            DEF_FREQUENCY_DOWN_DIFFERENTIAL,
  52        .ignore_nice = 0,
  53        .powersave_bias = 0,
  54};
  55
  56static void ondemand_powersave_bias_init_cpu(int cpu)
  57{
  58        struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
  59
  60        dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
  61        dbs_info->freq_lo = 0;
  62}
  63
  64/*
  65 * Not all CPUs want IO time to be accounted as busy; this depends on how
  66 * efficient idling at a higher frequency/voltage is.
  67 * Pavel Machek says this is not so for various generations of AMD and old
  68 * Intel systems.
  69 * Mike Chan (android.com) claims this is also not true for ARM.
  70 * Because of this, whitelist specific known (series) of CPUs by default, and
  71 * leave all others up to the user.
  72 */
  73static int should_io_be_busy(void)
  74{
  75#if defined(CONFIG_X86)
  76        /*
  77         * For Intel, Core 2 (model 15) and later have an efficient idle.
  78         */
  79        if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
  80                        boot_cpu_data.x86 == 6 &&
  81                        boot_cpu_data.x86_model >= 15)
  82                return 1;
  83#endif
  84        return 0;
  85}
  86
  87/*
  88 * Find right freq to be set now with powersave_bias on.
  89 * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
  90 * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
  91 */
  92static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
  93                unsigned int freq_next, unsigned int relation)
  94{
  95        unsigned int freq_req, freq_reduc, freq_avg;
  96        unsigned int freq_hi, freq_lo;
  97        unsigned int index = 0;
  98        unsigned int jiffies_total, jiffies_hi, jiffies_lo;
  99        struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
 100                                                   policy->cpu);
 101
 102        if (!dbs_info->freq_table) {
 103                dbs_info->freq_lo = 0;
 104                dbs_info->freq_lo_jiffies = 0;
 105                return freq_next;
 106        }
 107
 108        cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
 109                        relation, &index);
 110        freq_req = dbs_info->freq_table[index].frequency;
 111        freq_reduc = freq_req * od_tuners.powersave_bias / 1000;
 112        freq_avg = freq_req - freq_reduc;
 113
 114        /* Find freq bounds for freq_avg in freq_table */
 115        index = 0;
 116        cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 117                        CPUFREQ_RELATION_H, &index);
 118        freq_lo = dbs_info->freq_table[index].frequency;
 119        index = 0;
 120        cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 121                        CPUFREQ_RELATION_L, &index);
 122        freq_hi = dbs_info->freq_table[index].frequency;
 123
 124        /* Find out how long we have to be in hi and lo freqs */
 125        if (freq_hi == freq_lo) {
 126                dbs_info->freq_lo = 0;
 127                dbs_info->freq_lo_jiffies = 0;
 128                return freq_lo;
 129        }
 130        jiffies_total = usecs_to_jiffies(od_tuners.sampling_rate);
 131        jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
 132        jiffies_hi += ((freq_hi - freq_lo) / 2);
 133        jiffies_hi /= (freq_hi - freq_lo);
 134        jiffies_lo = jiffies_total - jiffies_hi;
 135        dbs_info->freq_lo = freq_lo;
 136        dbs_info->freq_lo_jiffies = jiffies_lo;
 137        dbs_info->freq_hi_jiffies = jiffies_hi;
 138        return freq_hi;
 139}
 140
 141static void ondemand_powersave_bias_init(void)
 142{
 143        int i;
 144        for_each_online_cpu(i) {
 145                ondemand_powersave_bias_init_cpu(i);
 146        }
 147}
 148
 149static void dbs_freq_increase(struct cpufreq_policy *p, unsigned int freq)
 150{
 151        if (od_tuners.powersave_bias)
 152                freq = powersave_bias_target(p, freq, CPUFREQ_RELATION_H);
 153        else if (p->cur == p->max)
 154                return;
 155
 156        __cpufreq_driver_target(p, freq, od_tuners.powersave_bias ?
 157                        CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
 158}
 159
 160/*
 161 * Every sampling_rate, we check, if current idle time is less than 20%
 162 * (default), then we try to increase frequency. Every sampling_rate, we look
 163 * for the lowest frequency which can sustain the load while keeping idle time
 164 * over 30%. If such a frequency exist, we try to decrease to this frequency.
 165 *
 166 * Any frequency increase takes it to the maximum frequency. Frequency reduction
 167 * happens at minimum steps of 5% (default) of current frequency
 168 */
 169static void od_check_cpu(int cpu, unsigned int load_freq)
 170{
 171        struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 172        struct cpufreq_policy *policy = dbs_info->cdbs.cur_policy;
 173
 174        dbs_info->freq_lo = 0;
 175
 176        /* Check for frequency increase */
 177        if (load_freq > od_tuners.up_threshold * policy->cur) {
 178                /* If switching to max speed, apply sampling_down_factor */
 179                if (policy->cur < policy->max)
 180                        dbs_info->rate_mult =
 181                                od_tuners.sampling_down_factor;
 182                dbs_freq_increase(policy, policy->max);
 183                return;
 184        }
 185
 186        /* Check for frequency decrease */
 187        /* if we cannot reduce the frequency anymore, break out early */
 188        if (policy->cur == policy->min)
 189                return;
 190
 191        /*
 192         * The optimal frequency is the frequency that is the lowest that can
 193         * support the current CPU usage without triggering the up policy. To be
 194         * safe, we focus 10 points under the threshold.
 195         */
 196        if (load_freq < od_tuners.adj_up_threshold * policy->cur) {
 197                unsigned int freq_next;
 198                freq_next = load_freq / od_tuners.adj_up_threshold;
 199
 200                /* No longer fully busy, reset rate_mult */
 201                dbs_info->rate_mult = 1;
 202
 203                if (freq_next < policy->min)
 204                        freq_next = policy->min;
 205
 206                if (!od_tuners.powersave_bias) {
 207                        __cpufreq_driver_target(policy, freq_next,
 208                                        CPUFREQ_RELATION_L);
 209                } else {
 210                        int freq = powersave_bias_target(policy, freq_next,
 211                                        CPUFREQ_RELATION_L);
 212                        __cpufreq_driver_target(policy, freq,
 213                                        CPUFREQ_RELATION_L);
 214                }
 215        }
 216}
 217
 218static void od_dbs_timer(struct work_struct *work)
 219{
 220        struct delayed_work *dw = to_delayed_work(work);
 221        struct od_cpu_dbs_info_s *dbs_info =
 222                container_of(work, struct od_cpu_dbs_info_s, cdbs.work.work);
 223        unsigned int cpu = dbs_info->cdbs.cur_policy->cpu;
 224        struct od_cpu_dbs_info_s *core_dbs_info = &per_cpu(od_cpu_dbs_info,
 225                        cpu);
 226        int delay, sample_type = core_dbs_info->sample_type;
 227        bool eval_load;
 228
 229        mutex_lock(&core_dbs_info->cdbs.timer_mutex);
 230        eval_load = need_load_eval(&core_dbs_info->cdbs,
 231                        od_tuners.sampling_rate);
 232
 233        /* Common NORMAL_SAMPLE setup */
 234        core_dbs_info->sample_type = OD_NORMAL_SAMPLE;
 235        if (sample_type == OD_SUB_SAMPLE) {
 236                delay = core_dbs_info->freq_lo_jiffies;
 237                if (eval_load)
 238                        __cpufreq_driver_target(core_dbs_info->cdbs.cur_policy,
 239                                                core_dbs_info->freq_lo,
 240                                                CPUFREQ_RELATION_H);
 241        } else {
 242                if (eval_load)
 243                        dbs_check_cpu(&od_dbs_data, cpu);
 244                if (core_dbs_info->freq_lo) {
 245                        /* Setup timer for SUB_SAMPLE */
 246                        core_dbs_info->sample_type = OD_SUB_SAMPLE;
 247                        delay = core_dbs_info->freq_hi_jiffies;
 248                } else {
 249                        delay = delay_for_sampling_rate(od_tuners.sampling_rate
 250                                                * core_dbs_info->rate_mult);
 251                }
 252        }
 253
 254        schedule_delayed_work_on(smp_processor_id(), dw, delay);
 255        mutex_unlock(&core_dbs_info->cdbs.timer_mutex);
 256}
 257
 258/************************** sysfs interface ************************/
 259
 260static ssize_t show_sampling_rate_min(struct kobject *kobj,
 261                                      struct attribute *attr, char *buf)
 262{
 263        return sprintf(buf, "%u\n", od_dbs_data.min_sampling_rate);
 264}
 265
 266/**
 267 * update_sampling_rate - update sampling rate effective immediately if needed.
 268 * @new_rate: new sampling rate
 269 *
 270 * If new rate is smaller than the old, simply updating
 271 * dbs_tuners_int.sampling_rate might not be appropriate. For example, if the
 272 * original sampling_rate was 1 second and the requested new sampling rate is 10
 273 * ms because the user needs immediate reaction from ondemand governor, but not
 274 * sure if higher frequency will be required or not, then, the governor may
 275 * change the sampling rate too late; up to 1 second later. Thus, if we are
 276 * reducing the sampling rate, we need to make the new value effective
 277 * immediately.
 278 */
 279static void update_sampling_rate(unsigned int new_rate)
 280{
 281        int cpu;
 282
 283        od_tuners.sampling_rate = new_rate = max(new_rate,
 284                        od_dbs_data.min_sampling_rate);
 285
 286        for_each_online_cpu(cpu) {
 287                struct cpufreq_policy *policy;
 288                struct od_cpu_dbs_info_s *dbs_info;
 289                unsigned long next_sampling, appointed_at;
 290
 291                policy = cpufreq_cpu_get(cpu);
 292                if (!policy)
 293                        continue;
 294                if (policy->governor != &cpufreq_gov_ondemand) {
 295                        cpufreq_cpu_put(policy);
 296                        continue;
 297                }
 298                dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 299                cpufreq_cpu_put(policy);
 300
 301                mutex_lock(&dbs_info->cdbs.timer_mutex);
 302
 303                if (!delayed_work_pending(&dbs_info->cdbs.work)) {
 304                        mutex_unlock(&dbs_info->cdbs.timer_mutex);
 305                        continue;
 306                }
 307
 308                next_sampling = jiffies + usecs_to_jiffies(new_rate);
 309                appointed_at = dbs_info->cdbs.work.timer.expires;
 310
 311                if (time_before(next_sampling, appointed_at)) {
 312
 313                        mutex_unlock(&dbs_info->cdbs.timer_mutex);
 314                        cancel_delayed_work_sync(&dbs_info->cdbs.work);
 315                        mutex_lock(&dbs_info->cdbs.timer_mutex);
 316
 317                        schedule_delayed_work_on(cpu, &dbs_info->cdbs.work,
 318                                        usecs_to_jiffies(new_rate));
 319
 320                }
 321                mutex_unlock(&dbs_info->cdbs.timer_mutex);
 322        }
 323}
 324
 325static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b,
 326                                   const char *buf, size_t count)
 327{
 328        unsigned int input;
 329        int ret;
 330        ret = sscanf(buf, "%u", &input);
 331        if (ret != 1)
 332                return -EINVAL;
 333        update_sampling_rate(input);
 334        return count;
 335}
 336
 337static ssize_t store_io_is_busy(struct kobject *a, struct attribute *b,
 338                                   const char *buf, size_t count)
 339{
 340        unsigned int input;
 341        int ret;
 342
 343        ret = sscanf(buf, "%u", &input);
 344        if (ret != 1)
 345                return -EINVAL;
 346        od_tuners.io_is_busy = !!input;
 347        return count;
 348}
 349
 350static ssize_t store_up_threshold(struct kobject *a, struct attribute *b,
 351                                  const char *buf, size_t count)
 352{
 353        unsigned int input;
 354        int ret;
 355        ret = sscanf(buf, "%u", &input);
 356
 357        if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 358                        input < MIN_FREQUENCY_UP_THRESHOLD) {
 359                return -EINVAL;
 360        }
 361        /* Calculate the new adj_up_threshold */
 362        od_tuners.adj_up_threshold += input;
 363        od_tuners.adj_up_threshold -= od_tuners.up_threshold;
 364
 365        od_tuners.up_threshold = input;
 366        return count;
 367}
 368
 369static ssize_t store_sampling_down_factor(struct kobject *a,
 370                        struct attribute *b, const char *buf, size_t count)
 371{
 372        unsigned int input, j;
 373        int ret;
 374        ret = sscanf(buf, "%u", &input);
 375
 376        if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
 377                return -EINVAL;
 378        od_tuners.sampling_down_factor = input;
 379
 380        /* Reset down sampling multiplier in case it was active */
 381        for_each_online_cpu(j) {
 382                struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
 383                                j);
 384                dbs_info->rate_mult = 1;
 385        }
 386        return count;
 387}
 388
 389static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b,
 390                                      const char *buf, size_t count)
 391{
 392        unsigned int input;
 393        int ret;
 394
 395        unsigned int j;
 396
 397        ret = sscanf(buf, "%u", &input);
 398        if (ret != 1)
 399                return -EINVAL;
 400
 401        if (input > 1)
 402                input = 1;
 403
 404        if (input == od_tuners.ignore_nice) { /* nothing to do */
 405                return count;
 406        }
 407        od_tuners.ignore_nice = input;
 408
 409        /* we need to re-evaluate prev_cpu_idle */
 410        for_each_online_cpu(j) {
 411                struct od_cpu_dbs_info_s *dbs_info;
 412                dbs_info = &per_cpu(od_cpu_dbs_info, j);
 413                dbs_info->cdbs.prev_cpu_idle = get_cpu_idle_time(j,
 414                                                &dbs_info->cdbs.prev_cpu_wall);
 415                if (od_tuners.ignore_nice)
 416                        dbs_info->cdbs.prev_cpu_nice =
 417                                kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 418
 419        }
 420        return count;
 421}
 422
 423static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b,
 424                                    const char *buf, size_t count)
 425{
 426        unsigned int input;
 427        int ret;
 428        ret = sscanf(buf, "%u", &input);
 429
 430        if (ret != 1)
 431                return -EINVAL;
 432
 433        if (input > 1000)
 434                input = 1000;
 435
 436        od_tuners.powersave_bias = input;
 437        ondemand_powersave_bias_init();
 438        return count;
 439}
 440
 441show_one(od, sampling_rate, sampling_rate);
 442show_one(od, io_is_busy, io_is_busy);
 443show_one(od, up_threshold, up_threshold);
 444show_one(od, sampling_down_factor, sampling_down_factor);
 445show_one(od, ignore_nice_load, ignore_nice);
 446show_one(od, powersave_bias, powersave_bias);
 447
 448define_one_global_rw(sampling_rate);
 449define_one_global_rw(io_is_busy);
 450define_one_global_rw(up_threshold);
 451define_one_global_rw(sampling_down_factor);
 452define_one_global_rw(ignore_nice_load);
 453define_one_global_rw(powersave_bias);
 454define_one_global_ro(sampling_rate_min);
 455
 456static struct attribute *dbs_attributes[] = {
 457        &sampling_rate_min.attr,
 458        &sampling_rate.attr,
 459        &up_threshold.attr,
 460        &sampling_down_factor.attr,
 461        &ignore_nice_load.attr,
 462        &powersave_bias.attr,
 463        &io_is_busy.attr,
 464        NULL
 465};
 466
 467static struct attribute_group od_attr_group = {
 468        .attrs = dbs_attributes,
 469        .name = "ondemand",
 470};
 471
 472/************************** sysfs end ************************/
 473
 474define_get_cpu_dbs_routines(od_cpu_dbs_info);
 475
 476static struct od_ops od_ops = {
 477        .io_busy = should_io_be_busy,
 478        .powersave_bias_init_cpu = ondemand_powersave_bias_init_cpu,
 479        .powersave_bias_target = powersave_bias_target,
 480        .freq_increase = dbs_freq_increase,
 481};
 482
 483static struct dbs_data od_dbs_data = {
 484        .governor = GOV_ONDEMAND,
 485        .attr_group = &od_attr_group,
 486        .tuners = &od_tuners,
 487        .get_cpu_cdbs = get_cpu_cdbs,
 488        .get_cpu_dbs_info_s = get_cpu_dbs_info_s,
 489        .gov_dbs_timer = od_dbs_timer,
 490        .gov_check_cpu = od_check_cpu,
 491        .gov_ops = &od_ops,
 492};
 493
 494static int od_cpufreq_governor_dbs(struct cpufreq_policy *policy,
 495                unsigned int event)
 496{
 497        return cpufreq_governor_dbs(&od_dbs_data, policy, event);
 498}
 499
 500#ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 501static
 502#endif
 503struct cpufreq_governor cpufreq_gov_ondemand = {
 504        .name                   = "ondemand",
 505        .governor               = od_cpufreq_governor_dbs,
 506        .max_transition_latency = TRANSITION_LATENCY_LIMIT,
 507        .owner                  = THIS_MODULE,
 508};
 509
 510static int __init cpufreq_gov_dbs_init(void)
 511{
 512        u64 idle_time;
 513        int cpu = get_cpu();
 514
 515        mutex_init(&od_dbs_data.mutex);
 516        idle_time = get_cpu_idle_time_us(cpu, NULL);
 517        put_cpu();
 518        if (idle_time != -1ULL) {
 519                /* Idle micro accounting is supported. Use finer thresholds */
 520                od_tuners.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
 521                od_tuners.adj_up_threshold = MICRO_FREQUENCY_UP_THRESHOLD -
 522                                             MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
 523                /*
 524                 * In nohz/micro accounting case we set the minimum frequency
 525                 * not depending on HZ, but fixed (very low). The deferred
 526                 * timer might skip some samples if idle/sleeping as needed.
 527                */
 528                od_dbs_data.min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
 529        } else {
 530                /* For correct statistics, we need 10 ticks for each measure */
 531                od_dbs_data.min_sampling_rate = MIN_SAMPLING_RATE_RATIO *
 532                        jiffies_to_usecs(10);
 533        }
 534
 535        return cpufreq_register_governor(&cpufreq_gov_ondemand);
 536}
 537
 538static void __exit cpufreq_gov_dbs_exit(void)
 539{
 540        cpufreq_unregister_governor(&cpufreq_gov_ondemand);
 541}
 542
 543MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 544MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 545MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 546        "Low Latency Frequency Transition capable processors");
 547MODULE_LICENSE("GPL");
 548
 549#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 550fs_initcall(cpufreq_gov_dbs_init);
 551#else
 552module_init(cpufreq_gov_dbs_init);
 553#endif
 554module_exit(cpufreq_gov_dbs_exit);
 555
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.