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         unsigned int cached_raw_freq;
  40
  41         /* The next fields are only needed if fast switch cannot be used. */
  42         struct irq_work irq_work;
  43         struct kthread_work work;
  44         struct mutex work_lock;
  45         struct kthread_worker worker;
  46         struct task_struct *thread;
  47         bool work_in_progress;
  48
  49         bool need_freq_update;
  50 };
  51
  52 struct sugov_cpu {
  53         struct update_util_data update_util;
  54         struct sugov_policy *sg_policy;
  55         unsigned int cpu;
  56
  57         bool iowait_boost_pending;
  58         unsigned int iowait_boost;
  59         unsigned int iowait_boost_max;
  60         u64 last_update;
  61
  62         /* The fields below are only needed when sharing a policy. */
  63         unsigned long util;
  64         unsigned long max;
  65         unsigned int flags;
  66
  67         /* The field below is for single-CPU policies only. */
  68 #ifdef CONFIG_NO_HZ_COMMON
  69         unsigned long saved_idle_calls;
  70 #endif
  71 };
  72
  73 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
  74
  75 /************************ Governor internals ***********************/
  76
  77 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
  78 {
  79         s64 delta_ns;
  80
  81         /*
  82          * Since cpufreq_update_util() is called with rq->lock held for
  83          * the @target_cpu, our per-cpu data is fully serialized.
  84          *
  85          * However, drivers cannot in general deal with cross-cpu
  86          * requests, so while get_next_freq() will work, our
  87          * sugov_update_commit() call may not for the fast switching platforms.
  88          *
  89          * Hence stop here for remote requests if they aren't supported
  90          * by the hardware, as calculating the frequency is pointless if
  91          * we cannot in fact act on it.
  92          *
  93          * For the slow switching platforms, the kthread is always scheduled on
  94          * the right set of CPUs and any CPU can find the next frequency and
  95          * schedule the kthread.
  96          */
  97         if (sg_policy->policy->fast_switch_enabled &&
  98             !cpufreq_can_do_remote_dvfs(sg_policy->policy))
  99                 return false;
 100
 101         if (sg_policy->work_in_progress)
 102                 return false;
 103
 104         if (unlikely(sg_policy->need_freq_update)) {
 105                 sg_policy->need_freq_update = false;
 106                 /*
 107                  * This happens when limits change, so forget the previous
 108                  * next_freq value and force an update.
 109                  */
 110                 sg_policy->next_freq = UINT_MAX;
 111                 return true;
 112         }
 113
 114         delta_ns = time - sg_policy->last_freq_update_time;
 115         return delta_ns >= sg_policy->freq_update_delay_ns;
 116 }
 117
 118 static void sugov_update_commit(struct sugov_policy *sg_policy, u64 time,
 119                                 unsigned int next_freq)
 120 {
 121         struct cpufreq_policy *policy = sg_policy->policy;
 122
 123         if (sg_policy->next_freq == next_freq)
 124                 return;
 125
 126         sg_policy->next_freq = next_freq;
 127         sg_policy->last_freq_update_time = time;
 128
 129         if (policy->fast_switch_enabled) {
 130                 next_freq = cpufreq_driver_fast_switch(policy, next_freq);
 131                 if (!next_freq)
 132                         return;
 133
 134                 policy->cur = next_freq;
 135                 trace_cpu_frequency(next_freq, smp_processor_id());
 136         } else {
 137                 sg_policy->work_in_progress = true;
 138                 irq_work_queue(&sg_policy->irq_work);
 139         }
 140 }
 141
 142 /**
 143  * get_next_freq - Compute a new frequency for a given cpufreq policy.
 144  * @sg_policy: schedutil policy object to compute the new frequency for.
 145  * @util: Current CPU utilization.
 146  * @max: CPU capacity.
 147  *
 148  * If the utilization is frequency-invariant, choose the new frequency to be
 149  * proportional to it, that is
 150  *
 151  * next_freq = C * max_freq * util / max
 152  *
 153  * Otherwise, approximate the would-be frequency-invariant utilization by
 154  * util_raw * (curr_freq / max_freq) which leads to
 155  *
 156  * next_freq = C * curr_freq * util_raw / max
 157  *
 158  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
 159  *
 160  * The lowest driver-supported frequency which is equal or greater than the raw
 161  * next_freq (as calculated above) is returned, subject to policy min/max and
 162  * cpufreq driver limitations.
 163  */
 164 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
 165                                   unsigned long util, unsigned long max)
 166 {
 167         struct cpufreq_policy *policy = sg_policy->policy;
 168         unsigned int freq = arch_scale_freq_invariant() ?
 169                                 policy->cpuinfo.max_freq : policy->cur;
 170
 171         freq = (freq + (freq >> 2)) * util / max;
 172
 173         if (freq == sg_policy->cached_raw_freq && sg_policy->next_freq != UINT_MAX)
 174                 return sg_policy->next_freq;
 175         sg_policy->cached_raw_freq = freq;
 176         return cpufreq_driver_resolve_freq(policy, freq);
 177 }
 178
 179 static void sugov_get_util(unsigned long *util, unsigned long *max, int cpu)
 180 {
 181         struct rq *rq = cpu_rq(cpu);
 182         unsigned long cfs_max;
 183
 184         cfs_max = arch_scale_cpu_capacity(NULL, cpu);
 185
 186         *util = min(rq->cfs.avg.util_avg, cfs_max);
 187         *max = cfs_max;
 188 }
 189
 190 static void sugov_set_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
 191                                    unsigned int flags)
 192 {
 193         if (flags & SCHED_CPUFREQ_IOWAIT) {
 194                 if (sg_cpu->iowait_boost_pending)
 195                         return;
 196
 197                 sg_cpu->iowait_boost_pending = true;
 198
 199                 if (sg_cpu->iowait_boost) {
 200                         sg_cpu->iowait_boost <<= 1;
 201                         if (sg_cpu->iowait_boost > sg_cpu->iowait_boost_max)
 202                                 sg_cpu->iowait_boost = sg_cpu->iowait_boost_max;
 203                 } else {
 204                         sg_cpu->iowait_boost = sg_cpu->sg_policy->policy->min;
 205                 }
 206         } else if (sg_cpu->iowait_boost) {
 207                 s64 delta_ns = time - sg_cpu->last_update;
 208
 209                 /* Clear iowait_boost if the CPU apprears to have been idle. */
 210                 if (delta_ns > TICK_NSEC) {
 211                         sg_cpu->iowait_boost = 0;
 212                         sg_cpu->iowait_boost_pending = false;
 213                 }
 214         }
 215 }
 216
 217 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, unsigned long *util,
 218                                unsigned long *max)
 219 {
 220         unsigned int boost_util, boost_max;
 221
 222         if (!sg_cpu->iowait_boost)
 223                 return;
 224
 225         if (sg_cpu->iowait_boost_pending) {
 226                 sg_cpu->iowait_boost_pending = false;
 227         } else {
 228                 sg_cpu->iowait_boost >>= 1;
 229                 if (sg_cpu->iowait_boost < sg_cpu->sg_policy->policy->min) {
 230                         sg_cpu->iowait_boost = 0;
 231                         return;
 232                 }
 233         }
 234
 235         boost_util = sg_cpu->iowait_boost;
 236         boost_max = sg_cpu->iowait_boost_max;
 237
 238         if (*util * boost_max < *max * boost_util) {
 239                 *util = boost_util;
 240                 *max = boost_max;
 241         }
 242 }
 243
 244 #ifdef CONFIG_NO_HZ_COMMON
 245 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
 246 {
 247         unsigned long idle_calls = tick_nohz_get_idle_calls();
 248         bool ret = idle_calls == sg_cpu->saved_idle_calls;
 249
 250         sg_cpu->saved_idle_calls = idle_calls;
 251         return ret;
 252 }
 253 #else
 254 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
 255 #endif /* CONFIG_NO_HZ_COMMON */
 256
 257 static void sugov_update_single(struct update_util_data *hook, u64 time,
 258                                 unsigned int flags)
 259 {
 260         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 261         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 262         struct cpufreq_policy *policy = sg_policy->policy;
 263         unsigned long util, max;
 264         unsigned int next_f;
 265         bool busy;
 266
 267         sugov_set_iowait_boost(sg_cpu, time, flags);
 268         sg_cpu->last_update = time;
 269
 270         if (!sugov_should_update_freq(sg_policy, time))
 271                 return;
 272
 273         busy = sugov_cpu_is_busy(sg_cpu);
 274
 275         if (flags & SCHED_CPUFREQ_RT_DL) {
 276                 next_f = policy->cpuinfo.max_freq;
 277         } else {
 278                 sugov_get_util(&util, &max, sg_cpu->cpu);
 279                 sugov_iowait_boost(sg_cpu, &util, &max);
 280                 next_f = get_next_freq(sg_policy, util, max);
 281                 /*
 282                  * Do not reduce the frequency if the CPU has not been idle
 283                  * recently, as the reduction is likely to be premature then.
 284                  */
 285                 if (busy && next_f < sg_policy->next_freq) {
 286                         next_f = sg_policy->next_freq;
 287
 288                         /* Reset cached freq as next_freq has changed */
 289                         sg_policy->cached_raw_freq = 0;
 290                 }
 291         }
 292         sugov_update_commit(sg_policy, time, next_f);
 293 }
 294
 295 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
 296 {
 297         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 298         struct cpufreq_policy *policy = sg_policy->policy;
 299         unsigned long util = 0, max = 1;
 300         unsigned int j;
 301
 302         for_each_cpu(j, policy->cpus) {
 303                 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
 304                 unsigned long j_util, j_max;
 305                 s64 delta_ns;
 306
 307                 /*
 308                  * If the CPU utilization was last updated before the previous
 309                  * frequency update and the time elapsed between the last update
 310                  * of the CPU utilization and the last frequency update is long
 311                  * enough, don't take the CPU into account as it probably is
 312                  * idle now (and clear iowait_boost for it).
 313                  */
 314                 delta_ns = time - j_sg_cpu->last_update;
 315                 if (delta_ns > TICK_NSEC) {
 316                         j_sg_cpu->iowait_boost = 0;
 317                         j_sg_cpu->iowait_boost_pending = false;
 318                         continue;
 319                 }
 320                 if (j_sg_cpu->flags & SCHED_CPUFREQ_RT_DL)
 321                         return policy->cpuinfo.max_freq;
 322
 323                 j_util = j_sg_cpu->util;
 324                 j_max = j_sg_cpu->max;
 325                 if (j_util * max > j_max * util) {
 326                         util = j_util;
 327                         max = j_max;
 328                 }
 329
 330                 sugov_iowait_boost(j_sg_cpu, &util, &max);
 331         }
 332
 333         return get_next_freq(sg_policy, util, max);
 334 }
 335
 336 static void sugov_update_shared(struct update_util_data *hook, u64 time,
 337                                 unsigned int flags)
 338 {
 339         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 340         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 341         unsigned long util, max;
 342         unsigned int next_f;
 343
 344         sugov_get_util(&util, &max, sg_cpu->cpu);
 345
 346         raw_spin_lock(&sg_policy->update_lock);
 347
 348         sg_cpu->util = util;
 349         sg_cpu->max = max;
 350         sg_cpu->flags = flags;
 351
 352         sugov_set_iowait_boost(sg_cpu, time, flags);
 353         sg_cpu->last_update = time;
 354
 355         if (sugov_should_update_freq(sg_policy, time)) {
 356                 if (flags & SCHED_CPUFREQ_RT_DL)
 357                         next_f = sg_policy->policy->cpuinfo.max_freq;
 358                 else
 359                         next_f = sugov_next_freq_shared(sg_cpu, time);
 360
 361                 sugov_update_commit(sg_policy, time, next_f);
 362         }
 363
 364         raw_spin_unlock(&sg_policy->update_lock);
 365 }
 366
 367 static void sugov_work(struct kthread_work *work)
 368 {
 369         struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
 370
 371         mutex_lock(&sg_policy->work_lock);
 372         __cpufreq_driver_target(sg_policy->policy, sg_policy->next_freq,
 373                                 CPUFREQ_RELATION_L);
 374         mutex_unlock(&sg_policy->work_lock);
 375
 376         sg_policy->work_in_progress = false;
 377 }
 378
 379 static void sugov_irq_work(struct irq_work *irq_work)
 380 {
 381         struct sugov_policy *sg_policy;
 382
 383         sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
 384
 385         /*
 386          * For RT and deadline tasks, the schedutil governor shoots the
 387          * frequency to maximum. Special care must be taken to ensure that this
 388          * kthread doesn't result in the same behavior.
 389          *
 390          * This is (mostly) guaranteed by the work_in_progress flag. The flag is
 391          * updated only at the end of the sugov_work() function and before that
 392          * the schedutil governor rejects all other frequency scaling requests.
 393          *
 394          * There is a very rare case though, where the RT thread yields right
 395          * after the work_in_progress flag is cleared. The effects of that are
 396          * neglected for now.
 397          */
 398         kthread_queue_work(&sg_policy->worker, &sg_policy->work);
 399 }
 400
 401 /************************** sysfs interface ************************/
 402
 403 static struct sugov_tunables *global_tunables;
 404 static DEFINE_MUTEX(global_tunables_lock);
 405
 406 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
 407 {
 408         return container_of(attr_set, struct sugov_tunables, attr_set);
 409 }
 410
 411 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
 412 {
 413         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 414
 415         return sprintf(buf, "%u\n", tunables->rate_limit_us);
 416 }
 417
 418 static ssize_t rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf,
 419                                    size_t count)
 420 {
 421         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 422         struct sugov_policy *sg_policy;
 423         unsigned int rate_limit_us;
 424
 425         if (kstrtouint(buf, 10, &rate_limit_us))
 426                 return -EINVAL;
 427
 428         tunables->rate_limit_us = rate_limit_us;
 429
 430         list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
 431                 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
 432
 433         return count;
 434 }
 435
 436 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
 437
 438 static struct attribute *sugov_attributes[] = {
 439         &rate_limit_us.attr,
 440         NULL
 441 };
 442
 443 static struct kobj_type sugov_tunables_ktype = {
 444         .default_attrs = sugov_attributes,
 445         .sysfs_ops = &governor_sysfs_ops,
 446 };
 447
 448 /********************** cpufreq governor interface *********************/
 449
 450 static struct cpufreq_governor schedutil_gov;
 451
 452 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
 453 {
 454         struct sugov_policy *sg_policy;
 455
 456         sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
 457         if (!sg_policy)
 458                 return NULL;
 459
 460         sg_policy->policy = policy;
 461         raw_spin_lock_init(&sg_policy->update_lock);
 462         return sg_policy;
 463 }
 464
 465 static void sugov_policy_free(struct sugov_policy *sg_policy)
 466 {
 467         kfree(sg_policy);
 468 }
 469
 470 static int sugov_kthread_create(struct sugov_policy *sg_policy)
 471 {
 472         struct task_struct *thread;
 473         struct sched_param param = { .sched_priority = MAX_USER_RT_PRIO / 2 };
 474         struct cpufreq_policy *policy = sg_policy->policy;
 475         int ret;
 476
 477         /* kthread only required for slow path */
 478         if (policy->fast_switch_enabled)
 479                 return 0;
 480
 481         kthread_init_work(&sg_policy->work, sugov_work);
 482         kthread_init_worker(&sg_policy->worker);
 483         thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
 484                                 "sugov:%d",
 485                                 cpumask_first(policy->related_cpus));
 486         if (IS_ERR(thread)) {
 487                 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
 488                 return PTR_ERR(thread);
 489         }
 490
 491         ret = sched_setscheduler_nocheck(thread, SCHED_FIFO, &param);
 492         if (ret) {
 493                 kthread_stop(thread);
 494                 pr_warn("%s: failed to set SCHED_FIFO\n", __func__);
 495                 return ret;
 496         }
 497
 498         sg_policy->thread = thread;
 499
 500         /* Kthread is bound to all CPUs by default */
 501         if (!policy->dvfs_possible_from_any_cpu)
 502                 kthread_bind_mask(thread, policy->related_cpus);
 503
 504         init_irq_work(&sg_policy->irq_work, sugov_irq_work);
 505         mutex_init(&sg_policy->work_lock);
 506
 507         wake_up_process(thread);
 508
 509         return 0;
 510 }
 511
 512 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
 513 {
 514         /* kthread only required for slow path */
 515         if (sg_policy->policy->fast_switch_enabled)
 516                 return;
 517
 518         kthread_flush_worker(&sg_policy->worker);
 519         kthread_stop(sg_policy->thread);
 520         mutex_destroy(&sg_policy->work_lock);
 521 }
 522
 523 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
 524 {
 525         struct sugov_tunables *tunables;
 526
 527         tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
 528         if (tunables) {
 529                 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
 530                 if (!have_governor_per_policy())
 531                         global_tunables = tunables;
 532         }
 533         return tunables;
 534 }
 535
 536 static void sugov_tunables_free(struct sugov_tunables *tunables)
 537 {
 538         if (!have_governor_per_policy())
 539                 global_tunables = NULL;
 540
 541         kfree(tunables);
 542 }
 543
 544 static int sugov_init(struct cpufreq_policy *policy)
 545 {
 546         struct sugov_policy *sg_policy;
 547         struct sugov_tunables *tunables;
 548         int ret = 0;
 549
 550         /* State should be equivalent to EXIT */
 551         if (policy->governor_data)
 552                 return -EBUSY;
 553
 554         cpufreq_enable_fast_switch(policy);
 555
 556         sg_policy = sugov_policy_alloc(policy);
 557         if (!sg_policy) {
 558                 ret = -ENOMEM;
 559                 goto disable_fast_switch;
 560         }
 561
 562         ret = sugov_kthread_create(sg_policy);
 563         if (ret)
 564                 goto free_sg_policy;
 565
 566         mutex_lock(&global_tunables_lock);
 567
 568         if (global_tunables) {
 569                 if (WARN_ON(have_governor_per_policy())) {
 570                         ret = -EINVAL;
 571                         goto stop_kthread;
 572                 }
 573                 policy->governor_data = sg_policy;
 574                 sg_policy->tunables = global_tunables;
 575
 576                 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
 577                 goto out;
 578         }
 579
 580         tunables = sugov_tunables_alloc(sg_policy);
 581         if (!tunables) {
 582                 ret = -ENOMEM;
 583                 goto stop_kthread;
 584         }
 585
 586         tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
 587
 588         policy->governor_data = sg_policy;
 589         sg_policy->tunables = tunables;
 590
 591         ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
 592                                    get_governor_parent_kobj(policy), "%s",
 593                                    schedutil_gov.name);
 594         if (ret)
 595                 goto fail;
 596
 597 out:
 598         mutex_unlock(&global_tunables_lock);
 599         return 0;
 600
 601 fail:
 602         policy->governor_data = NULL;
 603         sugov_tunables_free(tunables);
 604
 605 stop_kthread:
 606         sugov_kthread_stop(sg_policy);
 607
 608 free_sg_policy:
 609         mutex_unlock(&global_tunables_lock);
 610
 611         sugov_policy_free(sg_policy);
 612
 613 disable_fast_switch:
 614         cpufreq_disable_fast_switch(policy);
 615
 616         pr_err("initialization failed (error %d)\n", ret);
 617         return ret;
 618 }
 619
 620 static void sugov_exit(struct cpufreq_policy *policy)
 621 {
 622         struct sugov_policy *sg_policy = policy->governor_data;
 623         struct sugov_tunables *tunables = sg_policy->tunables;
 624         unsigned int count;
 625
 626         mutex_lock(&global_tunables_lock);
 627
 628         count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
 629         policy->governor_data = NULL;
 630         if (!count)
 631                 sugov_tunables_free(tunables);
 632
 633         mutex_unlock(&global_tunables_lock);
 634
 635         sugov_kthread_stop(sg_policy);
 636         sugov_policy_free(sg_policy);
 637         cpufreq_disable_fast_switch(policy);
 638 }
 639
 640 static int sugov_start(struct cpufreq_policy *policy)
 641 {
 642         struct sugov_policy *sg_policy = policy->governor_data;
 643         unsigned int cpu;
 644
 645         sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
 646         sg_policy->last_freq_update_time = 0;
 647         sg_policy->next_freq = UINT_MAX;
 648         sg_policy->work_in_progress = false;
 649         sg_policy->need_freq_update = false;
 650         sg_policy->cached_raw_freq = 0;
 651
 652         for_each_cpu(cpu, policy->cpus) {
 653                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 654
 655                 memset(sg_cpu, 0, sizeof(*sg_cpu));
 656                 sg_cpu->cpu = cpu;
 657                 sg_cpu->sg_policy = sg_policy;
 658                 sg_cpu->flags = SCHED_CPUFREQ_RT;
 659                 sg_cpu->iowait_boost_max = policy->cpuinfo.max_freq;
 660         }
 661
 662         for_each_cpu(cpu, policy->cpus) {
 663                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 664
 665                 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
 666                                              policy_is_shared(policy) ?
 667                                                         sugov_update_shared :
 668                                                         sugov_update_single);
 669         }
 670         return 0;
 671 }
 672
 673 static void sugov_stop(struct cpufreq_policy *policy)
 674 {
 675         struct sugov_policy *sg_policy = policy->governor_data;
 676         unsigned int cpu;
 677
 678         for_each_cpu(cpu, policy->cpus)
 679                 cpufreq_remove_update_util_hook(cpu);
 680
 681         synchronize_sched();
 682
 683         if (!policy->fast_switch_enabled) {
 684                 irq_work_sync(&sg_policy->irq_work);
 685                 kthread_cancel_work_sync(&sg_policy->work);
 686         }
 687 }
 688
 689 static void sugov_limits(struct cpufreq_policy *policy)
 690 {
 691         struct sugov_policy *sg_policy = policy->governor_data;
 692
 693         if (!policy->fast_switch_enabled) {
 694                 mutex_lock(&sg_policy->work_lock);
 695                 cpufreq_policy_apply_limits(policy);
 696                 mutex_unlock(&sg_policy->work_lock);
 697         }
 698
 699         sg_policy->need_freq_update = true;
 700 }
 701
 702 static struct cpufreq_governor schedutil_gov = {
 703         .name = "schedutil",
 704         .owner = THIS_MODULE,
 705         .dynamic_switching = true,
 706         .init = sugov_init,
 707         .exit = sugov_exit,
 708         .start = sugov_start,
 709         .stop = sugov_stop,
 710         .limits = sugov_limits,
 711 };
 712
 713 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
 714 struct cpufreq_governor *cpufreq_default_governor(void)
 715 {
 716         return &schedutil_gov;
 717 }
 718 #endif
 719
 720 static int __init sugov_register(void)
 721 {
 722         return cpufreq_register_governor(&schedutil_gov);
 723 }
 724 fs_initcall(sugov_register);