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_cpu(sg_cpu->cpu);
 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                     sg_policy->next_freq != UINT_MAX) {
 287                         next_f = sg_policy->next_freq;
 288
 289                         /* Reset cached freq as next_freq has changed */
 290                         sg_policy->cached_raw_freq = 0;
 291                 }
 292         }
 293         sugov_update_commit(sg_policy, time, next_f);
 294 }
 295
 296 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
 297 {
 298         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 299         struct cpufreq_policy *policy = sg_policy->policy;
 300         unsigned long util = 0, max = 1;
 301         unsigned int j;
 302
 303         for_each_cpu(j, policy->cpus) {
 304                 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
 305                 unsigned long j_util, j_max;
 306                 s64 delta_ns;
 307
 308                 /*
 309                  * If the CPU utilization was last updated before the previous
 310                  * frequency update and the time elapsed between the last update
 311                  * of the CPU utilization and the last frequency update is long
 312                  * enough, don't take the CPU into account as it probably is
 313                  * idle now (and clear iowait_boost for it).
 314                  */
 315                 delta_ns = time - j_sg_cpu->last_update;
 316                 if (delta_ns > TICK_NSEC) {
 317                         j_sg_cpu->iowait_boost = 0;
 318                         j_sg_cpu->iowait_boost_pending = false;
 319                         continue;
 320                 }
 321                 if (j_sg_cpu->flags & SCHED_CPUFREQ_RT_DL)
 322                         return policy->cpuinfo.max_freq;
 323
 324                 j_util = j_sg_cpu->util;
 325                 j_max = j_sg_cpu->max;
 326                 if (j_util * max > j_max * util) {
 327                         util = j_util;
 328                         max = j_max;
 329                 }
 330
 331                 sugov_iowait_boost(j_sg_cpu, &util, &max);
 332         }
 333
 334         return get_next_freq(sg_policy, util, max);
 335 }
 336
 337 static void sugov_update_shared(struct update_util_data *hook, u64 time,
 338                                 unsigned int flags)
 339 {
 340         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 341         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 342         unsigned long util, max;
 343         unsigned int next_f;
 344
 345         sugov_get_util(&util, &max, sg_cpu->cpu);
 346
 347         raw_spin_lock(&sg_policy->update_lock);
 348
 349         sg_cpu->util = util;
 350         sg_cpu->max = max;
 351         sg_cpu->flags = flags;
 352
 353         sugov_set_iowait_boost(sg_cpu, time, flags);
 354         sg_cpu->last_update = time;
 355
 356         if (sugov_should_update_freq(sg_policy, time)) {
 357                 if (flags & SCHED_CPUFREQ_RT_DL)
 358                         next_f = sg_policy->policy->cpuinfo.max_freq;
 359                 else
 360                         next_f = sugov_next_freq_shared(sg_cpu, time);
 361
 362                 sugov_update_commit(sg_policy, time, next_f);
 363         }
 364
 365         raw_spin_unlock(&sg_policy->update_lock);
 366 }
 367
 368 static void sugov_work(struct kthread_work *work)
 369 {
 370         struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
 371
 372         mutex_lock(&sg_policy->work_lock);
 373         __cpufreq_driver_target(sg_policy->policy, sg_policy->next_freq,
 374                                 CPUFREQ_RELATION_L);
 375         mutex_unlock(&sg_policy->work_lock);
 376
 377         sg_policy->work_in_progress = false;
 378 }
 379
 380 static void sugov_irq_work(struct irq_work *irq_work)
 381 {
 382         struct sugov_policy *sg_policy;
 383
 384         sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
 385
 386         /*
 387          * For RT and deadline tasks, the schedutil governor shoots the
 388          * frequency to maximum. Special care must be taken to ensure that this
 389          * kthread doesn't result in the same behavior.
 390          *
 391          * This is (mostly) guaranteed by the work_in_progress flag. The flag is
 392          * updated only at the end of the sugov_work() function and before that
 393          * the schedutil governor rejects all other frequency scaling requests.
 394          *
 395          * There is a very rare case though, where the RT thread yields right
 396          * after the work_in_progress flag is cleared. The effects of that are
 397          * neglected for now.
 398          */
 399         kthread_queue_work(&sg_policy->worker, &sg_policy->work);
 400 }
 401
 402 /************************** sysfs interface ************************/
 403
 404 static struct sugov_tunables *global_tunables;
 405 static DEFINE_MUTEX(global_tunables_lock);
 406
 407 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
 408 {
 409         return container_of(attr_set, struct sugov_tunables, attr_set);
 410 }
 411
 412 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
 413 {
 414         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 415
 416         return sprintf(buf, "%u\n", tunables->rate_limit_us);
 417 }
 418
 419 static ssize_t rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf,
 420                                    size_t count)
 421 {
 422         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 423         struct sugov_policy *sg_policy;
 424         unsigned int rate_limit_us;
 425
 426         if (kstrtouint(buf, 10, &rate_limit_us))
 427                 return -EINVAL;
 428
 429         tunables->rate_limit_us = rate_limit_us;
 430
 431         list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
 432                 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
 433
 434         return count;
 435 }
 436
 437 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
 438
 439 static struct attribute *sugov_attributes[] = {
 440         &rate_limit_us.attr,
 441         NULL
 442 };
 443
 444 static void sugov_tunables_free(struct kobject *kobj)
 445 {
 446         struct gov_attr_set *attr_set = container_of(kobj, struct gov_attr_set, kobj);
 447
 448         kfree(to_sugov_tunables(attr_set));
 449 }
 450
 451 static struct kobj_type sugov_tunables_ktype = {
 452         .default_attrs = sugov_attributes,
 453         .sysfs_ops = &governor_sysfs_ops,
 454         .release = &sugov_tunables_free,
 455 };
 456
 457 /********************** cpufreq governor interface *********************/
 458
 459 static struct cpufreq_governor schedutil_gov;
 460
 461 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
 462 {
 463         struct sugov_policy *sg_policy;
 464
 465         sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
 466         if (!sg_policy)
 467                 return NULL;
 468
 469         sg_policy->policy = policy;
 470         raw_spin_lock_init(&sg_policy->update_lock);
 471         return sg_policy;
 472 }
 473
 474 static void sugov_policy_free(struct sugov_policy *sg_policy)
 475 {
 476         kfree(sg_policy);
 477 }
 478
 479 static int sugov_kthread_create(struct sugov_policy *sg_policy)
 480 {
 481         struct task_struct *thread;
 482         struct sched_param param = { .sched_priority = MAX_USER_RT_PRIO / 2 };
 483         struct cpufreq_policy *policy = sg_policy->policy;
 484         int ret;
 485
 486         /* kthread only required for slow path */
 487         if (policy->fast_switch_enabled)
 488                 return 0;
 489
 490         kthread_init_work(&sg_policy->work, sugov_work);
 491         kthread_init_worker(&sg_policy->worker);
 492         thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
 493                                 "sugov:%d",
 494                                 cpumask_first(policy->related_cpus));
 495         if (IS_ERR(thread)) {
 496                 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
 497                 return PTR_ERR(thread);
 498         }
 499
 500         ret = sched_setscheduler_nocheck(thread, SCHED_FIFO, &param);
 501         if (ret) {
 502                 kthread_stop(thread);
 503                 pr_warn("%s: failed to set SCHED_FIFO\n", __func__);
 504                 return ret;
 505         }
 506
 507         sg_policy->thread = thread;
 508
 509         /* Kthread is bound to all CPUs by default */
 510         if (!policy->dvfs_possible_from_any_cpu)
 511                 kthread_bind_mask(thread, policy->related_cpus);
 512
 513         init_irq_work(&sg_policy->irq_work, sugov_irq_work);
 514         mutex_init(&sg_policy->work_lock);
 515
 516         wake_up_process(thread);
 517
 518         return 0;
 519 }
 520
 521 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
 522 {
 523         /* kthread only required for slow path */
 524         if (sg_policy->policy->fast_switch_enabled)
 525                 return;
 526
 527         kthread_flush_worker(&sg_policy->worker);
 528         kthread_stop(sg_policy->thread);
 529         mutex_destroy(&sg_policy->work_lock);
 530 }
 531
 532 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
 533 {
 534         struct sugov_tunables *tunables;
 535
 536         tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
 537         if (tunables) {
 538                 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
 539                 if (!have_governor_per_policy())
 540                         global_tunables = tunables;
 541         }
 542         return tunables;
 543 }
 544
 545 static void sugov_clear_global_tunables(void)
 546 {
 547         if (!have_governor_per_policy())
 548                 global_tunables = NULL;
 549 }
 550
 551 static int sugov_init(struct cpufreq_policy *policy)
 552 {
 553         struct sugov_policy *sg_policy;
 554         struct sugov_tunables *tunables;
 555         int ret = 0;
 556
 557         /* State should be equivalent to EXIT */
 558         if (policy->governor_data)
 559                 return -EBUSY;
 560
 561         cpufreq_enable_fast_switch(policy);
 562
 563         sg_policy = sugov_policy_alloc(policy);
 564         if (!sg_policy) {
 565                 ret = -ENOMEM;
 566                 goto disable_fast_switch;
 567         }
 568
 569         ret = sugov_kthread_create(sg_policy);
 570         if (ret)
 571                 goto free_sg_policy;
 572
 573         mutex_lock(&global_tunables_lock);
 574
 575         if (global_tunables) {
 576                 if (WARN_ON(have_governor_per_policy())) {
 577                         ret = -EINVAL;
 578                         goto stop_kthread;
 579                 }
 580                 policy->governor_data = sg_policy;
 581                 sg_policy->tunables = global_tunables;
 582
 583                 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
 584                 goto out;
 585         }
 586
 587         tunables = sugov_tunables_alloc(sg_policy);
 588         if (!tunables) {
 589                 ret = -ENOMEM;
 590                 goto stop_kthread;
 591         }
 592
 593         tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
 594
 595         policy->governor_data = sg_policy;
 596         sg_policy->tunables = tunables;
 597
 598         ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
 599                                    get_governor_parent_kobj(policy), "%s",
 600                                    schedutil_gov.name);
 601         if (ret)
 602                 goto fail;
 603
 604 out:
 605         mutex_unlock(&global_tunables_lock);
 606         return 0;
 607
 608 fail:
 609         kobject_put(&tunables->attr_set.kobj);
 610         policy->governor_data = NULL;
 611         sugov_clear_global_tunables();
 612
 613 stop_kthread:
 614         sugov_kthread_stop(sg_policy);
 615         mutex_unlock(&global_tunables_lock);
 616
 617 free_sg_policy:
 618         sugov_policy_free(sg_policy);
 619
 620 disable_fast_switch:
 621         cpufreq_disable_fast_switch(policy);
 622
 623         pr_err("initialization failed (error %d)\n", ret);
 624         return ret;
 625 }
 626
 627 static void sugov_exit(struct cpufreq_policy *policy)
 628 {
 629         struct sugov_policy *sg_policy = policy->governor_data;
 630         struct sugov_tunables *tunables = sg_policy->tunables;
 631         unsigned int count;
 632
 633         mutex_lock(&global_tunables_lock);
 634
 635         count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
 636         policy->governor_data = NULL;
 637         if (!count)
 638                 sugov_clear_global_tunables();
 639
 640         mutex_unlock(&global_tunables_lock);
 641
 642         sugov_kthread_stop(sg_policy);
 643         sugov_policy_free(sg_policy);
 644         cpufreq_disable_fast_switch(policy);
 645 }
 646
 647 static int sugov_start(struct cpufreq_policy *policy)
 648 {
 649         struct sugov_policy *sg_policy = policy->governor_data;
 650         unsigned int cpu;
 651
 652         sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
 653         sg_policy->last_freq_update_time = 0;
 654         sg_policy->next_freq = UINT_MAX;
 655         sg_policy->work_in_progress = false;
 656         sg_policy->need_freq_update = false;
 657         sg_policy->cached_raw_freq = 0;
 658
 659         for_each_cpu(cpu, policy->cpus) {
 660                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 661
 662                 memset(sg_cpu, 0, sizeof(*sg_cpu));
 663                 sg_cpu->cpu = cpu;
 664                 sg_cpu->sg_policy = sg_policy;
 665                 sg_cpu->flags = SCHED_CPUFREQ_RT;
 666                 sg_cpu->iowait_boost_max = policy->cpuinfo.max_freq;
 667         }
 668
 669         for_each_cpu(cpu, policy->cpus) {
 670                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 671
 672                 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
 673                                              policy_is_shared(policy) ?
 674                                                         sugov_update_shared :
 675                                                         sugov_update_single);
 676         }
 677         return 0;
 678 }
 679
 680 static void sugov_stop(struct cpufreq_policy *policy)
 681 {
 682         struct sugov_policy *sg_policy = policy->governor_data;
 683         unsigned int cpu;
 684
 685         for_each_cpu(cpu, policy->cpus)
 686                 cpufreq_remove_update_util_hook(cpu);
 687
 688         synchronize_sched();
 689
 690         if (!policy->fast_switch_enabled) {
 691                 irq_work_sync(&sg_policy->irq_work);
 692                 kthread_cancel_work_sync(&sg_policy->work);
 693         }
 694 }
 695
 696 static void sugov_limits(struct cpufreq_policy *policy)
 697 {
 698         struct sugov_policy *sg_policy = policy->governor_data;
 699
 700         if (!policy->fast_switch_enabled) {
 701                 mutex_lock(&sg_policy->work_lock);
 702                 cpufreq_policy_apply_limits(policy);
 703                 mutex_unlock(&sg_policy->work_lock);
 704         }
 705
 706         sg_policy->need_freq_update = true;
 707 }
 708
 709 static struct cpufreq_governor schedutil_gov = {
 710         .name = "schedutil",
 711         .owner = THIS_MODULE,
 712         .dynamic_switching = true,
 713         .init = sugov_init,
 714         .exit = sugov_exit,
 715         .start = sugov_start,
 716         .stop = sugov_stop,
 717         .limits = sugov_limits,
 718 };
 719
 720 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
 721 struct cpufreq_governor *cpufreq_default_governor(void)
 722 {
 723         return &schedutil_gov;
 724 }
 725 #endif
 726
 727 static int __init sugov_register(void)
 728 {
 729         return cpufreq_register_governor(&schedutil_gov);
 730 }
 731 fs_initcall(sugov_register);