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/cpu.h> 16#include <linux/percpu-defs.h> 17#include <linux/slab.h> 18#include <linux/tick.h> 19#include "cpufreq_governor.h" 20 21/* On-demand governor macros */ 22#define DEF_FREQUENCY_UP_THRESHOLD (80) 23#define DEF_SAMPLING_DOWN_FACTOR (1) 24#define MAX_SAMPLING_DOWN_FACTOR (100000) 25#define MICRO_FREQUENCY_UP_THRESHOLD (95) 26#define MICRO_FREQUENCY_MIN_SAMPLE_RATE (10000) 27#define MIN_FREQUENCY_UP_THRESHOLD (11) 28#define MAX_FREQUENCY_UP_THRESHOLD (100) 29 30static DEFINE_PER_CPU(struct od_cpu_dbs_info_s, od_cpu_dbs_info); 31 32static struct od_ops od_ops; 33 34#ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND 35static struct cpufreq_governor cpufreq_gov_ondemand; 36#endif 37 38static unsigned int default_powersave_bias; 39 40static void ondemand_powersave_bias_init_cpu(int cpu) 41{ 42 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu); 43 44 dbs_info->freq_table = cpufreq_frequency_get_table(cpu); 45 dbs_info->freq_lo = 0; 46} 47 48/* 49 * Not all CPUs want IO time to be accounted as busy; this depends on how 50 * efficient idling at a higher frequency/voltage is. 51 * Pavel Machek says this is not so for various generations of AMD and old 52 * Intel systems. 53 * Mike Chan (android.com) claims this is also not true for ARM. 54 * Because of this, whitelist specific known (series) of CPUs by default, and 55 * leave all others up to the user. 56 */ 57static int should_io_be_busy(void) 58{ 59#if defined(CONFIG_X86) 60 /* 61 * For Intel, Core 2 (model 15) and later have an efficient idle. 62 */ 63 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL && 64 boot_cpu_data.x86 == 6 && 65 boot_cpu_data.x86_model >= 15) 66 return 1; 67#endif 68 return 0; 69} 70 71/* 72 * Find right freq to be set now with powersave_bias on. 73 * Returns the freq_hi to be used right now and will set freq_hi_jiffies, 74 * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs. 75 */ 76static unsigned int generic_powersave_bias_target(struct cpufreq_policy *policy, 77 unsigned int freq_next, unsigned int relation) 78{ 79 unsigned int freq_req, freq_reduc, freq_avg; 80 unsigned int freq_hi, freq_lo; 81 unsigned int index = 0; 82 unsigned int jiffies_total, jiffies_hi, jiffies_lo; 83 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, 84 policy->cpu); 85 struct dbs_data *dbs_data = policy->governor_data; 86 struct od_dbs_tuners *od_tuners = dbs_data->tuners; 87 88 if (!dbs_info->freq_table) { 89 dbs_info->freq_lo = 0; 90 dbs_info->freq_lo_jiffies = 0; 91 return freq_next; 92 } 93 94 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next, 95 relation, &index); 96 freq_req = dbs_info->freq_table[index].frequency; 97 freq_reduc = freq_req * od_tuners->powersave_bias / 1000; 98 freq_avg = freq_req - freq_reduc; 99 100 /* Find freq bounds for freq_avg in freq_table */ 101 index = 0; 102 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg, 103 CPUFREQ_RELATION_H, &index); 104 freq_lo = dbs_info->freq_table[index].frequency; 105 index = 0; 106 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg, 107 CPUFREQ_RELATION_L, &index); 108 freq_hi = dbs_info->freq_table[index].frequency; 109 110 /* Find out how long we have to be in hi and lo freqs */ 111 if (freq_hi == freq_lo) { 112 dbs_info->freq_lo = 0; 113 dbs_info->freq_lo_jiffies = 0; 114 return freq_lo; 115 } 116 jiffies_total = usecs_to_jiffies(od_tuners->sampling_rate); 117 jiffies_hi = (freq_avg - freq_lo) * jiffies_total; 118 jiffies_hi += ((freq_hi - freq_lo) / 2); 119 jiffies_hi /= (freq_hi - freq_lo); 120 jiffies_lo = jiffies_total - jiffies_hi; 121 dbs_info->freq_lo = freq_lo; 122 dbs_info->freq_lo_jiffies = jiffies_lo; 123 dbs_info->freq_hi_jiffies = jiffies_hi; 124 return freq_hi; 125} 126 127static void ondemand_powersave_bias_init(void) 128{ 129 int i; 130 for_each_online_cpu(i) { 131 ondemand_powersave_bias_init_cpu(i); 132 } 133} 134 135static void dbs_freq_increase(struct cpufreq_policy *policy, unsigned int freq) 136{ 137 struct dbs_data *dbs_data = policy->governor_data; 138 struct od_dbs_tuners *od_tuners = dbs_data->tuners; 139 140 if (od_tuners->powersave_bias) 141 freq = od_ops.powersave_bias_target(policy, freq, 142 CPUFREQ_RELATION_H); 143 else if (policy->cur == policy->max) 144 return; 145 146 __cpufreq_driver_target(policy, freq, od_tuners->powersave_bias ? 147 CPUFREQ_RELATION_L : CPUFREQ_RELATION_H); 148} 149 150/* 151 * Every sampling_rate, we check, if current idle time is less than 20% 152 * (default), then we try to increase frequency. Else, we adjust the frequency 153 * proportional to load. 154 */ 155static void od_check_cpu(int cpu, unsigned int load) 156{ 157 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu); 158 struct cpufreq_policy *policy = dbs_info->cdbs.cur_policy; 159 struct dbs_data *dbs_data = policy->governor_data; 160 struct od_dbs_tuners *od_tuners = dbs_data->tuners; 161 162 dbs_info->freq_lo = 0; 163 164 /* Check for frequency increase */ 165 if (load > od_tuners->up_threshold) { 166 /* If switching to max speed, apply sampling_down_factor */ 167 if (policy->cur < policy->max) 168 dbs_info->rate_mult = 169 od_tuners->sampling_down_factor; 170 dbs_freq_increase(policy, policy->max); 171 } else { 172 /* Calculate the next frequency proportional to load */ 173 unsigned int freq_next, min_f, max_f; 174 175 min_f = policy->cpuinfo.min_freq; 176 max_f = policy->cpuinfo.max_freq; 177 freq_next = min_f + load * (max_f - min_f) / 100; 178 179 /* No longer fully busy, reset rate_mult */ 180 dbs_info->rate_mult = 1; 181 182 if (!od_tuners->powersave_bias) { 183 __cpufreq_driver_target(policy, freq_next, 184 CPUFREQ_RELATION_C); 185 return; 186 } 187 188 freq_next = od_ops.powersave_bias_target(policy, freq_next, 189 CPUFREQ_RELATION_L); 190 __cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_C); 191 } 192} 193 194static void od_dbs_timer(struct work_struct *work) 195{ 196 struct od_cpu_dbs_info_s *dbs_info = 197 container_of(work, struct od_cpu_dbs_info_s, cdbs.work.work); 198 unsigned int cpu = dbs_info->cdbs.cur_policy->cpu; 199 struct od_cpu_dbs_info_s *core_dbs_info = &per_cpu(od_cpu_dbs_info, 200 cpu); 201 struct dbs_data *dbs_data = dbs_info->cdbs.cur_policy->governor_data; 202 struct od_dbs_tuners *od_tuners = dbs_data->tuners; 203 int delay = 0, sample_type = core_dbs_info->sample_type; 204 bool modify_all = true; 205 206 mutex_lock(&core_dbs_info->cdbs.timer_mutex); 207 if (!need_load_eval(&core_dbs_info->cdbs, od_tuners->sampling_rate)) { 208 modify_all = false; 209 goto max_delay; 210 } 211 212 /* Common NORMAL_SAMPLE setup */ 213 core_dbs_info->sample_type = OD_NORMAL_SAMPLE; 214 if (sample_type == OD_SUB_SAMPLE) { 215 delay = core_dbs_info->freq_lo_jiffies; 216 __cpufreq_driver_target(core_dbs_info->cdbs.cur_policy, 217 core_dbs_info->freq_lo, CPUFREQ_RELATION_H); 218 } else { 219 dbs_check_cpu(dbs_data, cpu); 220 if (core_dbs_info->freq_lo) { 221 /* Setup timer for SUB_SAMPLE */ 222 core_dbs_info->sample_type = OD_SUB_SAMPLE; 223 delay = core_dbs_info->freq_hi_jiffies; 224 } 225 } 226 227max_delay: 228 if (!delay) 229 delay = delay_for_sampling_rate(od_tuners->sampling_rate 230 * core_dbs_info->rate_mult); 231 232 gov_queue_work(dbs_data, dbs_info->cdbs.cur_policy, delay, modify_all); 233 mutex_unlock(&core_dbs_info->cdbs.timer_mutex); 234} 235 236/************************** sysfs interface ************************/ 237static struct common_dbs_data od_dbs_cdata; 238 239/** 240 * update_sampling_rate - update sampling rate effective immediately if needed. 241 * @new_rate: new sampling rate 242 * 243 * If new rate is smaller than the old, simply updating 244 * dbs_tuners_int.sampling_rate might not be appropriate. For example, if the 245 * original sampling_rate was 1 second and the requested new sampling rate is 10 246 * ms because the user needs immediate reaction from ondemand governor, but not 247 * sure if higher frequency will be required or not, then, the governor may 248 * change the sampling rate too late; up to 1 second later. Thus, if we are 249 * reducing the sampling rate, we need to make the new value effective 250 * immediately. 251 */ 252static void update_sampling_rate(struct dbs_data *dbs_data, 253 unsigned int new_rate) 254{ 255 struct od_dbs_tuners *od_tuners = dbs_data->tuners; 256 int cpu; 257 258 od_tuners->sampling_rate = new_rate = max(new_rate, 259 dbs_data->min_sampling_rate); 260 261 for_each_online_cpu(cpu) { 262 struct cpufreq_policy *policy; 263 struct od_cpu_dbs_info_s *dbs_info; 264 unsigned long next_sampling, appointed_at; 265 266 policy = cpufreq_cpu_get(cpu); 267 if (!policy) 268 continue; 269 if (policy->governor != &cpufreq_gov_ondemand) { 270 cpufreq_cpu_put(policy); 271 continue; 272 } 273 dbs_info = &per_cpu(od_cpu_dbs_info, cpu); 274 cpufreq_cpu_put(policy); 275 276 mutex_lock(&dbs_info->cdbs.timer_mutex); 277 278 if (!delayed_work_pending(&dbs_info->cdbs.work)) { 279 mutex_unlock(&dbs_info->cdbs.timer_mutex); 280 continue; 281 } 282 283 next_sampling = jiffies + usecs_to_jiffies(new_rate); 284 appointed_at = dbs_info->cdbs.work.timer.expires; 285 286 if (time_before(next_sampling, appointed_at)) { 287 288 mutex_unlock(&dbs_info->cdbs.timer_mutex); 289 cancel_delayed_work_sync(&dbs_info->cdbs.work); 290 mutex_lock(&dbs_info->cdbs.timer_mutex); 291 292 gov_queue_work(dbs_data, dbs_info->cdbs.cur_policy, 293 usecs_to_jiffies(new_rate), true); 294 295 } 296 mutex_unlock(&dbs_info->cdbs.timer_mutex); 297 } 298} 299 300static ssize_t store_sampling_rate(struct dbs_data *dbs_data, const char *buf, 301 size_t count) 302{ 303 unsigned int input; 304 int ret; 305 ret = sscanf(buf, "%u", &input); 306 if (ret != 1) 307 return -EINVAL; 308 309 update_sampling_rate(dbs_data, input); 310 return count; 311} 312 313static ssize_t store_io_is_busy(struct dbs_data *dbs_data, const char *buf, 314 size_t count) 315{ 316 struct od_dbs_tuners *od_tuners = dbs_data->tuners; 317 unsigned int input; 318 int ret; 319 unsigned int j; 320 321 ret = sscanf(buf, "%u", &input); 322 if (ret != 1) 323 return -EINVAL; 324 od_tuners->io_is_busy = !!input; 325 326 /* we need to re-evaluate prev_cpu_idle */ 327 for_each_online_cpu(j) { 328 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, 329 j); 330 dbs_info->cdbs.prev_cpu_idle = get_cpu_idle_time(j, 331 &dbs_info->cdbs.prev_cpu_wall, od_tuners->io_is_busy); 332 } 333 return count; 334} 335 336static ssize_t store_up_threshold(struct dbs_data *dbs_data, const char *buf, 337 size_t count) 338{ 339 struct od_dbs_tuners *od_tuners = dbs_data->tuners; 340 unsigned int input; 341 int ret; 342 ret = sscanf(buf, "%u", &input); 343 344 if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD || 345 input < MIN_FREQUENCY_UP_THRESHOLD) { 346 return -EINVAL; 347 } 348 349 od_tuners->up_threshold = input; 350 return count; 351} 352 353static ssize_t store_sampling_down_factor(struct dbs_data *dbs_data, 354 const char *buf, size_t count) 355{ 356 struct od_dbs_tuners *od_tuners = dbs_data->tuners; 357 unsigned int input, j; 358 int ret; 359 ret = sscanf(buf, "%u", &input); 360 361 if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1) 362 return -EINVAL; 363 od_tuners->sampling_down_factor = input; 364 365 /* Reset down sampling multiplier in case it was active */ 366 for_each_online_cpu(j) { 367 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, 368 j); 369 dbs_info->rate_mult = 1; 370 } 371 return count; 372} 373 374static ssize_t store_ignore_nice_load(struct dbs_data *dbs_data, 375 const char *buf, size_t count) 376{ 377 struct od_dbs_tuners *od_tuners = dbs_data->tuners; 378 unsigned int input; 379 int ret; 380 381 unsigned int j; 382 383 ret = sscanf(buf, "%u", &input); 384 if (ret != 1) 385 return -EINVAL; 386 387 if (input > 1) 388 input = 1; 389 390 if (input == od_tuners->ignore_nice_load) { /* nothing to do */ 391 return count; 392 } 393 od_tuners->ignore_nice_load = input; 394 395 /* we need to re-evaluate prev_cpu_idle */ 396 for_each_online_cpu(j) { 397 struct od_cpu_dbs_info_s *dbs_info; 398 dbs_info = &per_cpu(od_cpu_dbs_info, j); 399 dbs_info->cdbs.prev_cpu_idle = get_cpu_idle_time(j, 400 &dbs_info->cdbs.prev_cpu_wall, od_tuners->io_is_busy); 401 if (od_tuners->ignore_nice_load) 402 dbs_info->cdbs.prev_cpu_nice = 403 kcpustat_cpu(j).cpustat[CPUTIME_NICE]; 404 405 } 406 return count; 407} 408 409static ssize_t store_powersave_bias(struct dbs_data *dbs_data, const char *buf, 410 size_t count) 411{ 412 struct od_dbs_tuners *od_tuners = dbs_data->tuners; 413 unsigned int input; 414 int ret; 415 ret = sscanf(buf, "%u", &input); 416 417 if (ret != 1) 418 return -EINVAL; 419 420 if (input > 1000) 421 input = 1000; 422 423 od_tuners->powersave_bias = input; 424 ondemand_powersave_bias_init(); 425 return count; 426} 427 428show_store_one(od, sampling_rate); 429show_store_one(od, io_is_busy); 430show_store_one(od, up_threshold); 431show_store_one(od, sampling_down_factor); 432show_store_one(od, ignore_nice_load); 433show_store_one(od, powersave_bias); 434declare_show_sampling_rate_min(od); 435 436gov_sys_pol_attr_rw(sampling_rate); 437gov_sys_pol_attr_rw(io_is_busy); 438gov_sys_pol_attr_rw(up_threshold); 439gov_sys_pol_attr_rw(sampling_down_factor); 440gov_sys_pol_attr_rw(ignore_nice_load); 441gov_sys_pol_attr_rw(powersave_bias); 442gov_sys_pol_attr_ro(sampling_rate_min); 443 444static struct attribute *dbs_attributes_gov_sys[] = { 445 &sampling_rate_min_gov_sys.attr, 446 &sampling_rate_gov_sys.attr, 447 &up_threshold_gov_sys.attr, 448 &sampling_down_factor_gov_sys.attr, 449 &ignore_nice_load_gov_sys.attr, 450 &powersave_bias_gov_sys.attr, 451 &io_is_busy_gov_sys.attr, 452 NULL 453}; 454 455static struct attribute_group od_attr_group_gov_sys = { 456 .attrs = dbs_attributes_gov_sys, 457 .name = "ondemand", 458}; 459 460static struct attribute *dbs_attributes_gov_pol[] = { 461 &sampling_rate_min_gov_pol.attr, 462 &sampling_rate_gov_pol.attr, 463 &up_threshold_gov_pol.attr, 464 &sampling_down_factor_gov_pol.attr, 465 &ignore_nice_load_gov_pol.attr, 466 &powersave_bias_gov_pol.attr, 467 &io_is_busy_gov_pol.attr, 468 NULL 469}; 470 471static struct attribute_group od_attr_group_gov_pol = { 472 .attrs = dbs_attributes_gov_pol, 473 .name = "ondemand", 474}; 475 476/************************** sysfs end ************************/ 477 478static int od_init(struct dbs_data *dbs_data) 479{ 480 struct od_dbs_tuners *tuners; 481 u64 idle_time; 482 int cpu; 483 484 tuners = kzalloc(sizeof(*tuners), GFP_KERNEL); 485 if (!tuners) { 486 pr_err("%s: kzalloc failed\n", __func__); 487 return -ENOMEM; 488 } 489 490 cpu = get_cpu(); 491 idle_time = get_cpu_idle_time_us(cpu, NULL); 492 put_cpu(); 493 if (idle_time != -1ULL) { 494 /* Idle micro accounting is supported. Use finer thresholds */ 495 tuners->up_threshold = MICRO_FREQUENCY_UP_THRESHOLD; 496 /* 497 * In nohz/micro accounting case we set the minimum frequency 498 * not depending on HZ, but fixed (very low). The deferred 499 * timer might skip some samples if idle/sleeping as needed. 500 */ 501 dbs_data->min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE; 502 } else { 503 tuners->up_threshold = DEF_FREQUENCY_UP_THRESHOLD; 504 505 /* For correct statistics, we need 10 ticks for each measure */ 506 dbs_data->min_sampling_rate = MIN_SAMPLING_RATE_RATIO * 507 jiffies_to_usecs(10); 508 } 509 510 tuners->sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR; 511 tuners->ignore_nice_load = 0; 512 tuners->powersave_bias = default_powersave_bias; 513 tuners->io_is_busy = should_io_be_busy(); 514 515 dbs_data->tuners = tuners; 516 mutex_init(&dbs_data->mutex); 517 return 0; 518} 519 520static void od_exit(struct dbs_data *dbs_data) 521{ 522 kfree(dbs_data->tuners); 523} 524 525define_get_cpu_dbs_routines(od_cpu_dbs_info); 526 527static struct od_ops od_ops = { 528 .powersave_bias_init_cpu = ondemand_powersave_bias_init_cpu, 529 .powersave_bias_target = generic_powersave_bias_target, 530 .freq_increase = dbs_freq_increase, 531}; 532 533static struct common_dbs_data od_dbs_cdata = { 534 .governor = GOV_ONDEMAND, 535 .attr_group_gov_sys = &od_attr_group_gov_sys, 536 .attr_group_gov_pol = &od_attr_group_gov_pol, 537 .get_cpu_cdbs = get_cpu_cdbs, 538 .get_cpu_dbs_info_s = get_cpu_dbs_info_s, 539 .gov_dbs_timer = od_dbs_timer, 540 .gov_check_cpu = od_check_cpu, 541 .gov_ops = &od_ops, 542 .init = od_init, 543 .exit = od_exit, 544}; 545 546static void od_set_powersave_bias(unsigned int powersave_bias) 547{ 548 struct cpufreq_policy *policy; 549 struct dbs_data *dbs_data; 550 struct od_dbs_tuners *od_tuners; 551 unsigned int cpu; 552 cpumask_t done; 553 554 default_powersave_bias = powersave_bias; 555 cpumask_clear(&done); 556 557 get_online_cpus(); 558 for_each_online_cpu(cpu) { 559 if (cpumask_test_cpu(cpu, &done)) 560 continue; 561 562 policy = per_cpu(od_cpu_dbs_info, cpu).cdbs.cur_policy; 563 if (!policy) 564 continue; 565 566 cpumask_or(&done, &done, policy->cpus); 567 568 if (policy->governor != &cpufreq_gov_ondemand) 569 continue; 570 571 dbs_data = policy->governor_data; 572 od_tuners = dbs_data->tuners; 573 od_tuners->powersave_bias = default_powersave_bias; 574 } 575 put_online_cpus(); 576} 577 578void od_register_powersave_bias_handler(unsigned int (*f) 579 (struct cpufreq_policy *, unsigned int, unsigned int), 580 unsigned int powersave_bias) 581{ 582 od_ops.powersave_bias_target = f; 583 od_set_powersave_bias(powersave_bias); 584} 585EXPORT_SYMBOL_GPL(od_register_powersave_bias_handler); 586 587void od_unregister_powersave_bias_handler(void) 588{ 589 od_ops.powersave_bias_target = generic_powersave_bias_target; 590 od_set_powersave_bias(0); 591} 592EXPORT_SYMBOL_GPL(od_unregister_powersave_bias_handler); 593 594static int od_cpufreq_governor_dbs(struct cpufreq_policy *policy, 595 unsigned int event) 596{ 597 return cpufreq_governor_dbs(policy, &od_dbs_cdata, event); 598} 599 600#ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND 601static 602#endif 603struct cpufreq_governor cpufreq_gov_ondemand = { 604 .name = "ondemand", 605 .governor = od_cpufreq_governor_dbs, 606 .max_transition_latency = TRANSITION_LATENCY_LIMIT, 607 .owner = THIS_MODULE, 608}; 609 610static int __init cpufreq_gov_dbs_init(void) 611{ 612 return cpufreq_register_governor(&cpufreq_gov_ondemand); 613} 614 615static void __exit cpufreq_gov_dbs_exit(void) 616{ 617 cpufreq_unregister_governor(&cpufreq_gov_ondemand); 618} 619 620MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>"); 621MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>"); 622MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for " 623 "Low Latency Frequency Transition capable processors"); 624MODULE_LICENSE("GPL"); 625 626#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND 627fs_initcall(cpufreq_gov_dbs_init); 628#else 629module_init(cpufreq_gov_dbs_init); 630#endif 631module_exit(cpufreq_gov_dbs_exit); 632