1 /* 2 * drivers/cpufreq/cpufreq_conservative.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 * (C) 2009 Alexander Clouter <alex@digriz.org.uk> 8 * 9 * This program is free software; you can redistribute it and/or modify 10 * it under the terms of the GNU General Public License version 2 as 11 * published by the Free Software Foundation. 12 */ 13 14 #include <linux/kernel.h> 15 #include <linux/module.h> 16 #include <linux/init.h> 17 #include <linux/cpufreq.h> 18 #include <linux/cpu.h> 19 #include <linux/jiffies.h> 20 #include <linux/kernel_stat.h> 21 #include <linux/mutex.h> 22 #include <linux/hrtimer.h> 23 #include <linux/tick.h> 24 #include <linux/ktime.h> 25 #include <linux/sched.h> 26 27 /* 28 * dbs is used in this file as a shortform for demandbased switching 29 * It helps to keep variable names smaller, simpler 30 */ 31 32 #define DEF_FREQUENCY_UP_THRESHOLD (80) 33 #define DEF_FREQUENCY_DOWN_THRESHOLD (20) 34 35 /* 36 * The polling frequency of this governor depends on the capability of 37 * the processor. Default polling frequency is 1000 times the transition 38 * latency of the processor. The governor will work on any processor with 39 * transition latency <= 10mS, using appropriate sampling 40 * rate. 41 * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL) 42 * this governor will not work. 43 * All times here are in uS. 44 */ 45 #define MIN_SAMPLING_RATE_RATIO (2) 46 47 static unsigned int min_sampling_rate; 48 49 #define LATENCY_MULTIPLIER (1000) 50 #define MIN_LATENCY_MULTIPLIER (100) 51 #define DEF_SAMPLING_DOWN_FACTOR (1) 52 #define MAX_SAMPLING_DOWN_FACTOR (10) 53 #define TRANSITION_LATENCY_LIMIT (10 * 1000 * 1000) 54 55 static void do_dbs_timer(struct work_struct *work); 56 57 struct cpu_dbs_info_s { 58 cputime64_t prev_cpu_idle; 59 cputime64_t prev_cpu_wall; 60 cputime64_t prev_cpu_nice; 61 struct cpufreq_policy *cur_policy; 62 struct delayed_work work; 63 unsigned int down_skip; 64 unsigned int requested_freq; 65 int cpu; 66 unsigned int enable:1; 67 /* 68 * percpu mutex that serializes governor limit change with 69 * do_dbs_timer invocation. We do not want do_dbs_timer to run 70 * when user is changing the governor or limits. 71 */ 72 struct mutex timer_mutex; 73 }; 74 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cs_cpu_dbs_info); 75 76 static unsigned int dbs_enable; /* number of CPUs using this policy */ 77 78 /* 79 * dbs_mutex protects dbs_enable in governor start/stop. 80 */ 81 static DEFINE_MUTEX(dbs_mutex); 82 83 static struct dbs_tuners { 84 unsigned int sampling_rate; 85 unsigned int sampling_down_factor; 86 unsigned int up_threshold; 87 unsigned int down_threshold; 88 unsigned int ignore_nice; 89 unsigned int freq_step; 90 } dbs_tuners_ins = { 91 .up_threshold = DEF_FREQUENCY_UP_THRESHOLD, 92 .down_threshold = DEF_FREQUENCY_DOWN_THRESHOLD, 93 .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR, 94 .ignore_nice = 0, 95 .freq_step = 5, 96 }; 97 98 static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu, 99 cputime64_t *wall) 100 { 101 cputime64_t idle_time; 102 cputime64_t cur_wall_time; 103 cputime64_t busy_time; 104 105 cur_wall_time = jiffies64_to_cputime64(get_jiffies_64()); 106 busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user, 107 kstat_cpu(cpu).cpustat.system); 108 109 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq); 110 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq); 111 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal); 112 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice); 113 114 idle_time = cputime64_sub(cur_wall_time, busy_time); 115 if (wall) 116 *wall = (cputime64_t)jiffies_to_usecs(cur_wall_time); 117 118 return (cputime64_t)jiffies_to_usecs(idle_time); 119 } 120 121 static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall) 122 { 123 u64 idle_time = get_cpu_idle_time_us(cpu, wall); 124 125 if (idle_time == -1ULL) 126 return get_cpu_idle_time_jiffy(cpu, wall); 127 128 return idle_time; 129 } 130 131 /* keep track of frequency transitions */ 132 static int 133 dbs_cpufreq_notifier(struct notifier_block *nb, unsigned long val, 134 void *data) 135 { 136 struct cpufreq_freqs *freq = data; 137 struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cs_cpu_dbs_info, 138 freq->cpu); 139 140 struct cpufreq_policy *policy; 141 142 if (!this_dbs_info->enable) 143 return 0; 144 145 policy = this_dbs_info->cur_policy; 146 147 /* 148 * we only care if our internally tracked freq moves outside 149 * the 'valid' ranges of freqency available to us otherwise 150 * we do not change it 151 */ 152 if (this_dbs_info->requested_freq > policy->max 153 || this_dbs_info->requested_freq < policy->min) 154 this_dbs_info->requested_freq = freq->new; 155 156 return 0; 157 } 158 159 static struct notifier_block dbs_cpufreq_notifier_block = { 160 .notifier_call = dbs_cpufreq_notifier 161 }; 162 163 /************************** sysfs interface ************************/ 164 static ssize_t show_sampling_rate_min(struct kobject *kobj, 165 struct attribute *attr, char *buf) 166 { 167 return sprintf(buf, "%u\n", min_sampling_rate); 168 } 169 170 define_one_global_ro(sampling_rate_min); 171 172 /* cpufreq_conservative Governor Tunables */ 173 #define show_one(file_name, object) \ 174 static ssize_t show_##file_name \ 175 (struct kobject *kobj, struct attribute *attr, char *buf) \ 176 { \ 177 return sprintf(buf, "%u\n", dbs_tuners_ins.object); \ 178 } 179 show_one(sampling_rate, sampling_rate); 180 show_one(sampling_down_factor, sampling_down_factor); 181 show_one(up_threshold, up_threshold); 182 show_one(down_threshold, down_threshold); 183 show_one(ignore_nice_load, ignore_nice); 184 show_one(freq_step, freq_step); 185 186 static ssize_t store_sampling_down_factor(struct kobject *a, 187 struct attribute *b, 188 const char *buf, size_t count) 189 { 190 unsigned int input; 191 int ret; 192 ret = sscanf(buf, "%u", &input); 193 194 if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1) 195 return -EINVAL; 196 197 dbs_tuners_ins.sampling_down_factor = input; 198 return count; 199 } 200 201 static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b, 202 const char *buf, size_t count) 203 { 204 unsigned int input; 205 int ret; 206 ret = sscanf(buf, "%u", &input); 207 208 if (ret != 1) 209 return -EINVAL; 210 211 dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate); 212 return count; 213 } 214 215 static ssize_t store_up_threshold(struct kobject *a, struct attribute *b, 216 const char *buf, size_t count) 217 { 218 unsigned int input; 219 int ret; 220 ret = sscanf(buf, "%u", &input); 221 222 if (ret != 1 || input > 100 || 223 input <= dbs_tuners_ins.down_threshold) 224 return -EINVAL; 225 226 dbs_tuners_ins.up_threshold = input; 227 return count; 228 } 229 230 static ssize_t store_down_threshold(struct kobject *a, struct attribute *b, 231 const char *buf, size_t count) 232 { 233 unsigned int input; 234 int ret; 235 ret = sscanf(buf, "%u", &input); 236 237 /* cannot be lower than 11 otherwise freq will not fall */ 238 if (ret != 1 || input < 11 || input > 100 || 239 input >= dbs_tuners_ins.up_threshold) 240 return -EINVAL; 241 242 dbs_tuners_ins.down_threshold = input; 243 return count; 244 } 245 246 static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b, 247 const char *buf, size_t count) 248 { 249 unsigned int input; 250 int ret; 251 252 unsigned int j; 253 254 ret = sscanf(buf, "%u", &input); 255 if (ret != 1) 256 return -EINVAL; 257 258 if (input > 1) 259 input = 1; 260 261 if (input == dbs_tuners_ins.ignore_nice) /* nothing to do */ 262 return count; 263 264 dbs_tuners_ins.ignore_nice = input; 265 266 /* we need to re-evaluate prev_cpu_idle */ 267 for_each_online_cpu(j) { 268 struct cpu_dbs_info_s *dbs_info; 269 dbs_info = &per_cpu(cs_cpu_dbs_info, j); 270 dbs_info->prev_cpu_idle = get_cpu_idle_time(j, 271 &dbs_info->prev_cpu_wall); 272 if (dbs_tuners_ins.ignore_nice) 273 dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice; 274 } 275 return count; 276 } 277 278 static ssize_t store_freq_step(struct kobject *a, struct attribute *b, 279 const char *buf, size_t count) 280 { 281 unsigned int input; 282 int ret; 283 ret = sscanf(buf, "%u", &input); 284 285 if (ret != 1) 286 return -EINVAL; 287 288 if (input > 100) 289 input = 100; 290 291 /* no need to test here if freq_step is zero as the user might actually 292 * want this, they would be crazy though :) */ 293 dbs_tuners_ins.freq_step = input; 294 return count; 295 } 296 297 define_one_global_rw(sampling_rate); 298 define_one_global_rw(sampling_down_factor); 299 define_one_global_rw(up_threshold); 300 define_one_global_rw(down_threshold); 301 define_one_global_rw(ignore_nice_load); 302 define_one_global_rw(freq_step); 303 304 static struct attribute *dbs_attributes[] = { 305 &sampling_rate_min.attr, 306 &sampling_rate.attr, 307 &sampling_down_factor.attr, 308 &up_threshold.attr, 309 &down_threshold.attr, 310 &ignore_nice_load.attr, 311 &freq_step.attr, 312 NULL 313 }; 314 315 static struct attribute_group dbs_attr_group = { 316 .attrs = dbs_attributes, 317 .name = "conservative", 318 }; 319 320 /************************** sysfs end ************************/ 321 322 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info) 323 { 324 unsigned int load = 0; 325 unsigned int max_load = 0; 326 unsigned int freq_target; 327 328 struct cpufreq_policy *policy; 329 unsigned int j; 330 331 policy = this_dbs_info->cur_policy; 332 333 /* 334 * Every sampling_rate, we check, if current idle time is less 335 * than 20% (default), then we try to increase frequency 336 * Every sampling_rate*sampling_down_factor, we check, if current 337 * idle time is more than 80%, then we try to decrease frequency 338 * 339 * Any frequency increase takes it to the maximum frequency. 340 * Frequency reduction happens at minimum steps of 341 * 5% (default) of maximum frequency 342 */ 343 344 /* Get Absolute Load */ 345 for_each_cpu(j, policy->cpus) { 346 struct cpu_dbs_info_s *j_dbs_info; 347 cputime64_t cur_wall_time, cur_idle_time; 348 unsigned int idle_time, wall_time; 349 350 j_dbs_info = &per_cpu(cs_cpu_dbs_info, j); 351 352 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time); 353 354 wall_time = (unsigned int) cputime64_sub(cur_wall_time, 355 j_dbs_info->prev_cpu_wall); 356 j_dbs_info->prev_cpu_wall = cur_wall_time; 357 358 idle_time = (unsigned int) cputime64_sub(cur_idle_time, 359 j_dbs_info->prev_cpu_idle); 360 j_dbs_info->prev_cpu_idle = cur_idle_time; 361 362 if (dbs_tuners_ins.ignore_nice) { 363 cputime64_t cur_nice; 364 unsigned long cur_nice_jiffies; 365 366 cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice, 367 j_dbs_info->prev_cpu_nice); 368 /* 369 * Assumption: nice time between sampling periods will 370 * be less than 2^32 jiffies for 32 bit sys 371 */ 372 cur_nice_jiffies = (unsigned long) 373 cputime64_to_jiffies64(cur_nice); 374 375 j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice; 376 idle_time += jiffies_to_usecs(cur_nice_jiffies); 377 } 378 379 if (unlikely(!wall_time || wall_time < idle_time)) 380 continue; 381 382 load = 100 * (wall_time - idle_time) / wall_time; 383 384 if (load > max_load) 385 max_load = load; 386 } 387 388 /* 389 * break out if we 'cannot' reduce the speed as the user might 390 * want freq_step to be zero 391 */ 392 if (dbs_tuners_ins.freq_step == 0) 393 return; 394 395 /* Check for frequency increase */ 396 if (max_load > dbs_tuners_ins.up_threshold) { 397 this_dbs_info->down_skip = 0; 398 399 /* if we are already at full speed then break out early */ 400 if (this_dbs_info->requested_freq == policy->max) 401 return; 402 403 freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100; 404 405 /* max freq cannot be less than 100. But who knows.... */ 406 if (unlikely(freq_target == 0)) 407 freq_target = 5; 408 409 this_dbs_info->requested_freq += freq_target; 410 if (this_dbs_info->requested_freq > policy->max) 411 this_dbs_info->requested_freq = policy->max; 412 413 __cpufreq_driver_target(policy, this_dbs_info->requested_freq, 414 CPUFREQ_RELATION_H); 415 return; 416 } 417 418 /* 419 * The optimal frequency is the frequency that is the lowest that 420 * can support the current CPU usage without triggering the up 421 * policy. To be safe, we focus 10 points under the threshold. 422 */ 423 if (max_load < (dbs_tuners_ins.down_threshold - 10)) { 424 freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100; 425 426 this_dbs_info->requested_freq -= freq_target; 427 if (this_dbs_info->requested_freq < policy->min) 428 this_dbs_info->requested_freq = policy->min; 429 430 /* 431 * if we cannot reduce the frequency anymore, break out early 432 */ 433 if (policy->cur == policy->min) 434 return; 435 436 __cpufreq_driver_target(policy, this_dbs_info->requested_freq, 437 CPUFREQ_RELATION_H); 438 return; 439 } 440 } 441 442 static void do_dbs_timer(struct work_struct *work) 443 { 444 struct cpu_dbs_info_s *dbs_info = 445 container_of(work, struct cpu_dbs_info_s, work.work); 446 unsigned int cpu = dbs_info->cpu; 447 448 /* We want all CPUs to do sampling nearly on same jiffy */ 449 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); 450 451 delay -= jiffies % delay; 452 453 mutex_lock(&dbs_info->timer_mutex); 454 455 dbs_check_cpu(dbs_info); 456 457 schedule_delayed_work_on(cpu, &dbs_info->work, delay); 458 mutex_unlock(&dbs_info->timer_mutex); 459 } 460 461 static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info) 462 { 463 /* We want all CPUs to do sampling nearly on same jiffy */ 464 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); 465 delay -= jiffies % delay; 466 467 dbs_info->enable = 1; 468 INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer); 469 schedule_delayed_work_on(dbs_info->cpu, &dbs_info->work, delay); 470 } 471 472 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info) 473 { 474 dbs_info->enable = 0; 475 cancel_delayed_work_sync(&dbs_info->work); 476 } 477 478 static int cpufreq_governor_dbs(struct cpufreq_policy *policy, 479 unsigned int event) 480 { 481 unsigned int cpu = policy->cpu; 482 struct cpu_dbs_info_s *this_dbs_info; 483 unsigned int j; 484 int rc; 485 486 this_dbs_info = &per_cpu(cs_cpu_dbs_info, cpu); 487 488 switch (event) { 489 case CPUFREQ_GOV_START: 490 if ((!cpu_online(cpu)) || (!policy->cur)) 491 return -EINVAL; 492 493 mutex_lock(&dbs_mutex); 494 495 for_each_cpu(j, policy->cpus) { 496 struct cpu_dbs_info_s *j_dbs_info; 497 j_dbs_info = &per_cpu(cs_cpu_dbs_info, j); 498 j_dbs_info->cur_policy = policy; 499 500 j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j, 501 &j_dbs_info->prev_cpu_wall); 502 if (dbs_tuners_ins.ignore_nice) { 503 j_dbs_info->prev_cpu_nice = 504 kstat_cpu(j).cpustat.nice; 505 } 506 } 507 this_dbs_info->down_skip = 0; 508 this_dbs_info->requested_freq = policy->cur; 509 510 mutex_init(&this_dbs_info->timer_mutex); 511 dbs_enable++; 512 /* 513 * Start the timerschedule work, when this governor 514 * is used for first time 515 */ 516 if (dbs_enable == 1) { 517 unsigned int latency; 518 /* policy latency is in nS. Convert it to uS first */ 519 latency = policy->cpuinfo.transition_latency / 1000; 520 if (latency == 0) 521 latency = 1; 522 523 rc = sysfs_create_group(cpufreq_global_kobject, 524 &dbs_attr_group); 525 if (rc) { 526 mutex_unlock(&dbs_mutex); 527 return rc; 528 } 529 530 /* 531 * conservative does not implement micro like ondemand 532 * governor, thus we are bound to jiffes/HZ 533 */ 534 min_sampling_rate = 535 MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10); 536 /* Bring kernel and HW constraints together */ 537 min_sampling_rate = max(min_sampling_rate, 538 MIN_LATENCY_MULTIPLIER * latency); 539 dbs_tuners_ins.sampling_rate = 540 max(min_sampling_rate, 541 latency * LATENCY_MULTIPLIER); 542 543 cpufreq_register_notifier( 544 &dbs_cpufreq_notifier_block, 545 CPUFREQ_TRANSITION_NOTIFIER); 546 } 547 mutex_unlock(&dbs_mutex); 548 549 dbs_timer_init(this_dbs_info); 550 551 break; 552 553 case CPUFREQ_GOV_STOP: 554 dbs_timer_exit(this_dbs_info); 555 556 mutex_lock(&dbs_mutex); 557 dbs_enable--; 558 mutex_destroy(&this_dbs_info->timer_mutex); 559 560 /* 561 * Stop the timerschedule work, when this governor 562 * is used for first time 563 */ 564 if (dbs_enable == 0) 565 cpufreq_unregister_notifier( 566 &dbs_cpufreq_notifier_block, 567 CPUFREQ_TRANSITION_NOTIFIER); 568 569 mutex_unlock(&dbs_mutex); 570 if (!dbs_enable) 571 sysfs_remove_group(cpufreq_global_kobject, 572 &dbs_attr_group); 573 574 break; 575 576 case CPUFREQ_GOV_LIMITS: 577 mutex_lock(&this_dbs_info->timer_mutex); 578 if (policy->max < this_dbs_info->cur_policy->cur) 579 __cpufreq_driver_target( 580 this_dbs_info->cur_policy, 581 policy->max, CPUFREQ_RELATION_H); 582 else if (policy->min > this_dbs_info->cur_policy->cur) 583 __cpufreq_driver_target( 584 this_dbs_info->cur_policy, 585 policy->min, CPUFREQ_RELATION_L); 586 mutex_unlock(&this_dbs_info->timer_mutex); 587 588 break; 589 } 590 return 0; 591 } 592 593 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE 594 static 595 #endif 596 struct cpufreq_governor cpufreq_gov_conservative = { 597 .name = "conservative", 598 .governor = cpufreq_governor_dbs, 599 .max_transition_latency = TRANSITION_LATENCY_LIMIT, 600 .owner = THIS_MODULE, 601 }; 602 603 static int __init cpufreq_gov_dbs_init(void) 604 { 605 return cpufreq_register_governor(&cpufreq_gov_conservative); 606 } 607 608 static void __exit cpufreq_gov_dbs_exit(void) 609 { 610 cpufreq_unregister_governor(&cpufreq_gov_conservative); 611 } 612 613 614 MODULE_AUTHOR("Alexander Clouter <alex@digriz.org.uk>"); 615 MODULE_DESCRIPTION("'cpufreq_conservative' - A dynamic cpufreq governor for " 616 "Low Latency Frequency Transition capable processors " 617 "optimised for use in a battery environment"); 618 MODULE_LICENSE("GPL"); 619 620 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE 621 fs_initcall(cpufreq_gov_dbs_init); 622 #else 623 module_init(cpufreq_gov_dbs_init); 624 #endif 625 module_exit(cpufreq_gov_dbs_exit); 626