kernel/sched/cpufreq_schedutil.c

   1 /*
   2  * CPUFreq governor based on scheduler-provided CPU utilization data.
   3  *
   4  * Copyright (C) 2016, Intel Corporation
   5  * Author: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
   6  *
   7  * This program is free software; you can redistribute it and/or modify
   8  * it under the terms of the GNU General Public License version 2 as
   9  * published by the Free Software Foundation.
  10  */
  11
  12 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  13
  14 #include <linux/cpufreq.h>
  15 #include <linux/kthread.h>
  16 #include <uapi/linux/sched/types.h>
  17 #include <linux/slab.h>
  18 #include <trace/events/power.h>
  19
  20 #include "sched.h"
  21
  22 #define SUGOV_KTHREAD_PRIORITY  50
  23
  24 struct sugov_tunables {
  25         struct gov_attr_set attr_set;
  26         unsigned int rate_limit_us;
  27 };
  28
  29 struct sugov_policy {
  30         struct cpufreq_policy *policy;
  31
  32         struct sugov_tunables *tunables;
  33         struct list_head tunables_hook;
  34
  35         raw_spinlock_t update_lock;  /* For shared policies */
  36         u64 last_freq_update_time;
  37         s64 freq_update_delay_ns;
  38         unsigned int next_freq;
  39
  40         /* The next fields are only needed if fast switch cannot be used. */
  41         struct irq_work irq_work;
  42         struct kthread_work work;
  43         struct mutex work_lock;
  44         struct kthread_worker worker;
  45         struct task_struct *thread;
  46         bool work_in_progress;
  47
  48         bool need_freq_update;
  49 };
  50
  51 struct sugov_cpu {
  52         struct update_util_data update_util;
  53         struct sugov_policy *sg_policy;
  54
  55         unsigned int cached_raw_freq;
  56         unsigned long iowait_boost;
  57         unsigned long iowait_boost_max;
  58         u64 last_update;
  59
  60         /* The fields below are only needed when sharing a policy. */
  61         unsigned long util;
  62         unsigned long max;
  63         unsigned int flags;
  64 };
  65
  66 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
  67
  68 /************************ Governor internals ***********************/
  69
  70 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
  71 {
  72         s64 delta_ns;
  73
  74         if (sg_policy->work_in_progress)
  75                 return false;
  76
  77         if (unlikely(sg_policy->need_freq_update)) {
  78                 sg_policy->need_freq_update = false;
  79                 /*
  80                  * This happens when limits change, so forget the previous
  81                  * next_freq value and force an update.
  82                  */
  83                 sg_policy->next_freq = UINT_MAX;
  84                 return true;
  85         }
  86
  87         delta_ns = time - sg_policy->last_freq_update_time;
  88         return delta_ns >= sg_policy->freq_update_delay_ns;
  89 }
  90
  91 static void sugov_update_commit(struct sugov_policy *sg_policy, u64 time,
  92                                 unsigned int next_freq)
  93 {
  94         struct cpufreq_policy *policy = sg_policy->policy;
  95
  96         sg_policy->last_freq_update_time = time;
  97
  98         if (policy->fast_switch_enabled) {
  99                 if (sg_policy->next_freq == next_freq) {
 100                         trace_cpu_frequency(policy->cur, smp_processor_id());
 101                         return;
 102                 }
 103                 sg_policy->next_freq = next_freq;
 104                 next_freq = cpufreq_driver_fast_switch(policy, next_freq);
 105                 if (next_freq == CPUFREQ_ENTRY_INVALID)
 106                         return;
 107
 108                 policy->cur = next_freq;
 109                 trace_cpu_frequency(next_freq, smp_processor_id());
 110         } else if (sg_policy->next_freq != next_freq) {
 111                 sg_policy->next_freq = next_freq;
 112                 sg_policy->work_in_progress = true;
 113                 irq_work_queue(&sg_policy->irq_work);
 114         }
 115 }
 116
 117 /**
 118  * get_next_freq - Compute a new frequency for a given cpufreq policy.
 119  * @sg_cpu: schedutil cpu object to compute the new frequency for.
 120  * @util: Current CPU utilization.
 121  * @max: CPU capacity.
 122  *
 123  * If the utilization is frequency-invariant, choose the new frequency to be
 124  * proportional to it, that is
 125  *
 126  * next_freq = C * max_freq * util / max
 127  *
 128  * Otherwise, approximate the would-be frequency-invariant utilization by
 129  * util_raw * (curr_freq / max_freq) which leads to
 130  *
 131  * next_freq = C * curr_freq * util_raw / max
 132  *
 133  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
 134  *
 135  * The lowest driver-supported frequency which is equal or greater than the raw
 136  * next_freq (as calculated above) is returned, subject to policy min/max and
 137  * cpufreq driver limitations.
 138  */
 139 static unsigned int get_next_freq(struct sugov_cpu *sg_cpu, unsigned long util,
 140                                   unsigned long max)
 141 {
 142         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 143         struct cpufreq_policy *policy = sg_policy->policy;
 144         unsigned int freq = arch_scale_freq_invariant() ?
 145                                 policy->cpuinfo.max_freq : policy->cur;
 146
 147         freq = (freq + (freq >> 2)) * util / max;
 148
 149         if (freq == sg_cpu->cached_raw_freq && sg_policy->next_freq != UINT_MAX)
 150                 return sg_policy->next_freq;
 151         sg_cpu->cached_raw_freq = freq;
 152         return cpufreq_driver_resolve_freq(policy, freq);
 153 }
 154
 155 static void sugov_get_util(unsigned long *util, unsigned long *max)
 156 {
 157         struct rq *rq = this_rq();
 158         unsigned long cfs_max;
 159
 160         cfs_max = arch_scale_cpu_capacity(NULL, smp_processor_id());
 161
 162         *util = min(rq->cfs.avg.util_avg, cfs_max);
 163         *max = cfs_max;
 164 }
 165
 166 static void sugov_set_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
 167                                    unsigned int flags)
 168 {
 169         if (flags & SCHED_CPUFREQ_IOWAIT) {
 170                 sg_cpu->iowait_boost = sg_cpu->iowait_boost_max;
 171         } else if (sg_cpu->iowait_boost) {
 172                 s64 delta_ns = time - sg_cpu->last_update;
 173
 174                 /* Clear iowait_boost if the CPU apprears to have been idle. */
 175                 if (delta_ns > TICK_NSEC)
 176                         sg_cpu->iowait_boost = 0;
 177         }
 178 }
 179
 180 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, unsigned long *util,
 181                                unsigned long *max)
 182 {
 183         unsigned long boost_util = sg_cpu->iowait_boost;
 184         unsigned long boost_max = sg_cpu->iowait_boost_max;
 185
 186         if (!boost_util)
 187                 return;
 188
 189         if (*util * boost_max < *max * boost_util) {
 190                 *util = boost_util;
 191                 *max = boost_max;
 192         }
 193         sg_cpu->iowait_boost >>= 1;
 194 }
 195
 196 static void sugov_update_single(struct update_util_data *hook, u64 time,
 197                                 unsigned int flags)
 198 {
 199         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 200         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 201         struct cpufreq_policy *policy = sg_policy->policy;
 202         unsigned long util, max;
 203         unsigned int next_f;
 204
 205         sugov_set_iowait_boost(sg_cpu, time, flags);
 206         sg_cpu->last_update = time;
 207
 208         if (!sugov_should_update_freq(sg_policy, time))
 209                 return;
 210
 211         if (flags & SCHED_CPUFREQ_RT_DL) {
 212                 next_f = policy->cpuinfo.max_freq;
 213         } else {
 214                 sugov_get_util(&util, &max);
 215                 sugov_iowait_boost(sg_cpu, &util, &max);
 216                 next_f = get_next_freq(sg_cpu, util, max);
 217         }
 218         sugov_update_commit(sg_policy, time, next_f);
 219 }
 220
 221 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu,
 222                                            unsigned long util, unsigned long max,
 223                                            unsigned int flags)
 224 {
 225         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 226         struct cpufreq_policy *policy = sg_policy->policy;
 227         unsigned int max_f = policy->cpuinfo.max_freq;
 228         u64 last_freq_update_time = sg_policy->last_freq_update_time;
 229         unsigned int j;
 230
 231         if (flags & SCHED_CPUFREQ_RT_DL)
 232                 return max_f;
 233
 234         sugov_iowait_boost(sg_cpu, &util, &max);
 235
 236         for_each_cpu(j, policy->cpus) {
 237                 struct sugov_cpu *j_sg_cpu;
 238                 unsigned long j_util, j_max;
 239                 s64 delta_ns;
 240
 241                 if (j == smp_processor_id())
 242                         continue;
 243
 244                 j_sg_cpu = &per_cpu(sugov_cpu, j);
 245                 /*
 246                  * If the CPU utilization was last updated before the previous
 247                  * frequency update and the time elapsed between the last update
 248                  * of the CPU utilization and the last frequency update is long
 249                  * enough, don't take the CPU into account as it probably is
 250                  * idle now (and clear iowait_boost for it).
 251                  */
 252                 delta_ns = last_freq_update_time - j_sg_cpu->last_update;
 253                 if (delta_ns > TICK_NSEC) {
 254                         j_sg_cpu->iowait_boost = 0;
 255                         continue;
 256                 }
 257                 if (j_sg_cpu->flags & SCHED_CPUFREQ_RT_DL)
 258                         return max_f;
 259
 260                 j_util = j_sg_cpu->util;
 261                 j_max = j_sg_cpu->max;
 262                 if (j_util * max > j_max * util) {
 263                         util = j_util;
 264                         max = j_max;
 265                 }
 266
 267                 sugov_iowait_boost(j_sg_cpu, &util, &max);
 268         }
 269
 270         return get_next_freq(sg_cpu, util, max);
 271 }
 272
 273 static void sugov_update_shared(struct update_util_data *hook, u64 time,
 274                                 unsigned int flags)
 275 {
 276         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 277         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 278         unsigned long util, max;
 279         unsigned int next_f;
 280
 281         sugov_get_util(&util, &max);
 282
 283         raw_spin_lock(&sg_policy->update_lock);
 284
 285         sg_cpu->util = util;
 286         sg_cpu->max = max;
 287         sg_cpu->flags = flags;
 288
 289         sugov_set_iowait_boost(sg_cpu, time, flags);
 290         sg_cpu->last_update = time;
 291
 292         if (sugov_should_update_freq(sg_policy, time)) {
 293                 next_f = sugov_next_freq_shared(sg_cpu, util, max, flags);
 294                 sugov_update_commit(sg_policy, time, next_f);
 295         }
 296
 297         raw_spin_unlock(&sg_policy->update_lock);
 298 }
 299
 300 static void sugov_work(struct kthread_work *work)
 301 {
 302         struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
 303
 304         mutex_lock(&sg_policy->work_lock);
 305         __cpufreq_driver_target(sg_policy->policy, sg_policy->next_freq,
 306                                 CPUFREQ_RELATION_L);
 307         mutex_unlock(&sg_policy->work_lock);
 308
 309         sg_policy->work_in_progress = false;
 310 }
 311
 312 static void sugov_irq_work(struct irq_work *irq_work)
 313 {
 314         struct sugov_policy *sg_policy;
 315
 316         sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
 317
 318         /*
 319          * For RT and deadline tasks, the schedutil governor shoots the
 320          * frequency to maximum. Special care must be taken to ensure that this
 321          * kthread doesn't result in the same behavior.
 322          *
 323          * This is (mostly) guaranteed by the work_in_progress flag. The flag is
 324          * updated only at the end of the sugov_work() function and before that
 325          * the schedutil governor rejects all other frequency scaling requests.
 326          *
 327          * There is a very rare case though, where the RT thread yields right
 328          * after the work_in_progress flag is cleared. The effects of that are
 329          * neglected for now.
 330          */
 331         kthread_queue_work(&sg_policy->worker, &sg_policy->work);
 332 }
 333
 334 /************************** sysfs interface ************************/
 335
 336 static struct sugov_tunables *global_tunables;
 337 static DEFINE_MUTEX(global_tunables_lock);
 338
 339 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
 340 {
 341         return container_of(attr_set, struct sugov_tunables, attr_set);
 342 }
 343
 344 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
 345 {
 346         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 347
 348         return sprintf(buf, "%u\n", tunables->rate_limit_us);
 349 }
 350
 351 static ssize_t rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf,
 352                                    size_t count)
 353 {
 354         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 355         struct sugov_policy *sg_policy;
 356         unsigned int rate_limit_us;
 357
 358         if (kstrtouint(buf, 10, &rate_limit_us))
 359                 return -EINVAL;
 360
 361         tunables->rate_limit_us = rate_limit_us;
 362
 363         list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
 364                 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
 365
 366         return count;
 367 }
 368
 369 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
 370
 371 static struct attribute *sugov_attributes[] = {
 372         &rate_limit_us.attr,
 373         NULL
 374 };
 375
 376 static struct kobj_type sugov_tunables_ktype = {
 377         .default_attrs = sugov_attributes,
 378         .sysfs_ops = &governor_sysfs_ops,
 379 };
 380
 381 /********************** cpufreq governor interface *********************/
 382
 383 static struct cpufreq_governor schedutil_gov;
 384
 385 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
 386 {
 387         struct sugov_policy *sg_policy;
 388
 389         sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
 390         if (!sg_policy)
 391                 return NULL;
 392
 393         sg_policy->policy = policy;
 394         raw_spin_lock_init(&sg_policy->update_lock);
 395         return sg_policy;
 396 }
 397
 398 static void sugov_policy_free(struct sugov_policy *sg_policy)
 399 {
 400         kfree(sg_policy);
 401 }
 402
 403 static int sugov_kthread_create(struct sugov_policy *sg_policy)
 404 {
 405         struct task_struct *thread;
 406         struct sched_param param = { .sched_priority = MAX_USER_RT_PRIO / 2 };
 407         struct cpufreq_policy *policy = sg_policy->policy;
 408         int ret;
 409
 410         /* kthread only required for slow path */
 411         if (policy->fast_switch_enabled)
 412                 return 0;
 413
 414         kthread_init_work(&sg_policy->work, sugov_work);
 415         kthread_init_worker(&sg_policy->worker);
 416         thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
 417                                 "sugov:%d",
 418                                 cpumask_first(policy->related_cpus));
 419         if (IS_ERR(thread)) {
 420                 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
 421                 return PTR_ERR(thread);
 422         }
 423
 424         ret = sched_setscheduler_nocheck(thread, SCHED_FIFO, &param);
 425         if (ret) {
 426                 kthread_stop(thread);
 427                 pr_warn("%s: failed to set SCHED_FIFO\n", __func__);
 428                 return ret;
 429         }
 430
 431         sg_policy->thread = thread;
 432         kthread_bind_mask(thread, policy->related_cpus);
 433         init_irq_work(&sg_policy->irq_work, sugov_irq_work);
 434         mutex_init(&sg_policy->work_lock);
 435
 436         wake_up_process(thread);
 437
 438         return 0;
 439 }
 440
 441 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
 442 {
 443         /* kthread only required for slow path */
 444         if (sg_policy->policy->fast_switch_enabled)
 445                 return;
 446
 447         kthread_flush_worker(&sg_policy->worker);
 448         kthread_stop(sg_policy->thread);
 449         mutex_destroy(&sg_policy->work_lock);
 450 }
 451
 452 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
 453 {
 454         struct sugov_tunables *tunables;
 455
 456         tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
 457         if (tunables) {
 458                 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
 459                 if (!have_governor_per_policy())
 460                         global_tunables = tunables;
 461         }
 462         return tunables;
 463 }
 464
 465 static void sugov_tunables_free(struct sugov_tunables *tunables)
 466 {
 467         if (!have_governor_per_policy())
 468                 global_tunables = NULL;
 469
 470         kfree(tunables);
 471 }
 472
 473 static int sugov_init(struct cpufreq_policy *policy)
 474 {
 475         struct sugov_policy *sg_policy;
 476         struct sugov_tunables *tunables;
 477         unsigned int lat;
 478         int ret = 0;
 479
 480         /* State should be equivalent to EXIT */
 481         if (policy->governor_data)
 482                 return -EBUSY;
 483
 484         cpufreq_enable_fast_switch(policy);
 485
 486         sg_policy = sugov_policy_alloc(policy);
 487         if (!sg_policy) {
 488                 ret = -ENOMEM;
 489                 goto disable_fast_switch;
 490         }
 491
 492         ret = sugov_kthread_create(sg_policy);
 493         if (ret)
 494                 goto free_sg_policy;
 495
 496         mutex_lock(&global_tunables_lock);
 497
 498         if (global_tunables) {
 499                 if (WARN_ON(have_governor_per_policy())) {
 500                         ret = -EINVAL;
 501                         goto stop_kthread;
 502                 }
 503                 policy->governor_data = sg_policy;
 504                 sg_policy->tunables = global_tunables;
 505
 506                 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
 507                 goto out;
 508         }
 509
 510         tunables = sugov_tunables_alloc(sg_policy);
 511         if (!tunables) {
 512                 ret = -ENOMEM;
 513                 goto stop_kthread;
 514         }
 515
 516         tunables->rate_limit_us = LATENCY_MULTIPLIER;
 517         lat = policy->cpuinfo.transition_latency / NSEC_PER_USEC;
 518         if (lat)
 519                 tunables->rate_limit_us *= lat;
 520
 521         policy->governor_data = sg_policy;
 522         sg_policy->tunables = tunables;
 523
 524         ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
 525                                    get_governor_parent_kobj(policy), "%s",
 526                                    schedutil_gov.name);
 527         if (ret)
 528                 goto fail;
 529
 530 out:
 531         mutex_unlock(&global_tunables_lock);
 532         return 0;
 533
 534 fail:
 535         policy->governor_data = NULL;
 536         sugov_tunables_free(tunables);
 537
 538 stop_kthread:
 539         sugov_kthread_stop(sg_policy);
 540
 541 free_sg_policy:
 542         mutex_unlock(&global_tunables_lock);
 543
 544         sugov_policy_free(sg_policy);
 545
 546 disable_fast_switch:
 547         cpufreq_disable_fast_switch(policy);
 548
 549         pr_err("initialization failed (error %d)\n", ret);
 550         return ret;
 551 }
 552
 553 static void sugov_exit(struct cpufreq_policy *policy)
 554 {
 555         struct sugov_policy *sg_policy = policy->governor_data;
 556         struct sugov_tunables *tunables = sg_policy->tunables;
 557         unsigned int count;
 558
 559         mutex_lock(&global_tunables_lock);
 560
 561         count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
 562         policy->governor_data = NULL;
 563         if (!count)
 564                 sugov_tunables_free(tunables);
 565
 566         mutex_unlock(&global_tunables_lock);
 567
 568         sugov_kthread_stop(sg_policy);
 569         sugov_policy_free(sg_policy);
 570         cpufreq_disable_fast_switch(policy);
 571 }
 572
 573 static int sugov_start(struct cpufreq_policy *policy)
 574 {
 575         struct sugov_policy *sg_policy = policy->governor_data;
 576         unsigned int cpu;
 577
 578         sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
 579         sg_policy->last_freq_update_time = 0;
 580         sg_policy->next_freq = UINT_MAX;
 581         sg_policy->work_in_progress = false;
 582         sg_policy->need_freq_update = false;
 583
 584         for_each_cpu(cpu, policy->cpus) {
 585                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 586
 587                 sg_cpu->sg_policy = sg_policy;
 588                 if (policy_is_shared(policy)) {
 589                         sg_cpu->util = 0;
 590                         sg_cpu->max = 0;
 591                         sg_cpu->flags = SCHED_CPUFREQ_RT;
 592                         sg_cpu->last_update = 0;
 593                         sg_cpu->cached_raw_freq = 0;
 594                         sg_cpu->iowait_boost = 0;
 595                         sg_cpu->iowait_boost_max = policy->cpuinfo.max_freq;
 596                         cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
 597                                                      sugov_update_shared);
 598                 } else {
 599                         cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
 600                                                      sugov_update_single);
 601                 }
 602         }
 603         return 0;
 604 }
 605
 606 static void sugov_stop(struct cpufreq_policy *policy)
 607 {
 608         struct sugov_policy *sg_policy = policy->governor_data;
 609         unsigned int cpu;
 610
 611         for_each_cpu(cpu, policy->cpus)
 612                 cpufreq_remove_update_util_hook(cpu);
 613
 614         synchronize_sched();
 615
 616         if (!policy->fast_switch_enabled) {
 617                 irq_work_sync(&sg_policy->irq_work);
 618                 kthread_cancel_work_sync(&sg_policy->work);
 619         }
 620 }
 621
 622 static void sugov_limits(struct cpufreq_policy *policy)
 623 {
 624         struct sugov_policy *sg_policy = policy->governor_data;
 625
 626         if (!policy->fast_switch_enabled) {
 627                 mutex_lock(&sg_policy->work_lock);
 628                 cpufreq_policy_apply_limits(policy);
 629                 mutex_unlock(&sg_policy->work_lock);
 630         }
 631
 632         sg_policy->need_freq_update = true;
 633 }
 634
 635 static struct cpufreq_governor schedutil_gov = {
 636         .name = "schedutil",
 637         .owner = THIS_MODULE,
 638         .init = sugov_init,
 639         .exit = sugov_exit,
 640         .start = sugov_start,
 641         .stop = sugov_stop,
 642         .limits = sugov_limits,
 643 };
 644
 645 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
 646 struct cpufreq_governor *cpufreq_default_governor(void)
 647 {
 648         return &schedutil_gov;
 649 }
 650 #endif
 651
 652 static int __init sugov_register(void)
 653 {
 654         return cpufreq_register_governor(&schedutil_gov);
 655 }
 656 fs_initcall(sugov_register);