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_cfs;
  64         unsigned long util_dl;
  65         unsigned long max;
  66         unsigned int flags;
  67
  68         /* The field below is for single-CPU policies only. */
  69 #ifdef CONFIG_NO_HZ_COMMON
  70         unsigned long saved_idle_calls;
  71 #endif
  72 };
  73
  74 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
  75
  76 /************************ Governor internals ***********************/
  77
  78 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
  79 {
  80         s64 delta_ns;
  81
  82         /*
  83          * Since cpufreq_update_util() is called with rq->lock held for
  84          * the @target_cpu, our per-cpu data is fully serialized.
  85          *
  86          * However, drivers cannot in general deal with cross-cpu
  87          * requests, so while get_next_freq() will work, our
  88          * sugov_update_commit() call may not for the fast switching platforms.
  89          *
  90          * Hence stop here for remote requests if they aren't supported
  91          * by the hardware, as calculating the frequency is pointless if
  92          * we cannot in fact act on it.
  93          *
  94          * For the slow switching platforms, the kthread is always scheduled on
  95          * the right set of CPUs and any CPU can find the next frequency and
  96          * schedule the kthread.
  97          */
  98         if (sg_policy->policy->fast_switch_enabled &&
  99             !cpufreq_can_do_remote_dvfs(sg_policy->policy))
 100                 return false;
 101
 102         if (sg_policy->work_in_progress)
 103                 return false;
 104
 105         if (unlikely(sg_policy->need_freq_update)) {
 106                 sg_policy->need_freq_update = false;
 107                 /*
 108                  * This happens when limits change, so forget the previous
 109                  * next_freq value and force an update.
 110                  */
 111                 sg_policy->next_freq = UINT_MAX;
 112                 return true;
 113         }
 114
 115         delta_ns = time - sg_policy->last_freq_update_time;
 116         return delta_ns >= sg_policy->freq_update_delay_ns;
 117 }
 118
 119 static void sugov_update_commit(struct sugov_policy *sg_policy, u64 time,
 120                                 unsigned int next_freq)
 121 {
 122         struct cpufreq_policy *policy = sg_policy->policy;
 123
 124         if (sg_policy->next_freq == next_freq)
 125                 return;
 126
 127         sg_policy->next_freq = next_freq;
 128         sg_policy->last_freq_update_time = time;
 129
 130         if (policy->fast_switch_enabled) {
 131                 next_freq = cpufreq_driver_fast_switch(policy, next_freq);
 132                 if (!next_freq)
 133                         return;
 134
 135                 policy->cur = next_freq;
 136                 trace_cpu_frequency(next_freq, smp_processor_id());
 137         } else {
 138                 sg_policy->work_in_progress = true;
 139                 irq_work_queue(&sg_policy->irq_work);
 140         }
 141 }
 142
 143 /**
 144  * get_next_freq - Compute a new frequency for a given cpufreq policy.
 145  * @sg_policy: schedutil policy object to compute the new frequency for.
 146  * @util: Current CPU utilization.
 147  * @max: CPU capacity.
 148  *
 149  * If the utilization is frequency-invariant, choose the new frequency to be
 150  * proportional to it, that is
 151  *
 152  * next_freq = C * max_freq * util / max
 153  *
 154  * Otherwise, approximate the would-be frequency-invariant utilization by
 155  * util_raw * (curr_freq / max_freq) which leads to
 156  *
 157  * next_freq = C * curr_freq * util_raw / max
 158  *
 159  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
 160  *
 161  * The lowest driver-supported frequency which is equal or greater than the raw
 162  * next_freq (as calculated above) is returned, subject to policy min/max and
 163  * cpufreq driver limitations.
 164  */
 165 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
 166                                   unsigned long util, unsigned long max)
 167 {
 168         struct cpufreq_policy *policy = sg_policy->policy;
 169         unsigned int freq = arch_scale_freq_invariant() ?
 170                                 policy->cpuinfo.max_freq : policy->cur;
 171
 172         freq = (freq + (freq >> 2)) * util / max;
 173
 174         if (freq == sg_policy->cached_raw_freq && sg_policy->next_freq != UINT_MAX)
 175                 return sg_policy->next_freq;
 176         sg_policy->cached_raw_freq = freq;
 177         return cpufreq_driver_resolve_freq(policy, freq);
 178 }
 179
 180 static void sugov_get_util(struct sugov_cpu *sg_cpu)
 181 {
 182         struct rq *rq = cpu_rq(sg_cpu->cpu);
 183
 184         sg_cpu->max = arch_scale_cpu_capacity(NULL, sg_cpu->cpu);
 185         sg_cpu->util_cfs = cpu_util_cfs(rq);
 186         sg_cpu->util_dl  = cpu_util_dl(rq);
 187 }
 188
 189 static unsigned long sugov_aggregate_util(struct sugov_cpu *sg_cpu)
 190 {
 191         /*
 192          * Ideally we would like to set util_dl as min/guaranteed freq and
 193          * util_cfs + util_dl as requested freq. However, cpufreq is not yet
 194          * ready for such an interface. So, we only do the latter for now.
 195          */
 196         return min(sg_cpu->util_cfs + sg_cpu->util_dl, sg_cpu->max);
 197 }
 198
 199 static void sugov_set_iowait_boost(struct sugov_cpu *sg_cpu, u64 time)
 200 {
 201         if (sg_cpu->flags & SCHED_CPUFREQ_IOWAIT) {
 202                 if (sg_cpu->iowait_boost_pending)
 203                         return;
 204
 205                 sg_cpu->iowait_boost_pending = true;
 206
 207                 if (sg_cpu->iowait_boost) {
 208                         sg_cpu->iowait_boost <<= 1;
 209                         if (sg_cpu->iowait_boost > sg_cpu->iowait_boost_max)
 210                                 sg_cpu->iowait_boost = sg_cpu->iowait_boost_max;
 211                 } else {
 212                         sg_cpu->iowait_boost = sg_cpu->sg_policy->policy->min;
 213                 }
 214         } else if (sg_cpu->iowait_boost) {
 215                 s64 delta_ns = time - sg_cpu->last_update;
 216
 217                 /* Clear iowait_boost if the CPU apprears to have been idle. */
 218                 if (delta_ns > TICK_NSEC) {
 219                         sg_cpu->iowait_boost = 0;
 220                         sg_cpu->iowait_boost_pending = false;
 221                 }
 222         }
 223 }
 224
 225 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, unsigned long *util,
 226                                unsigned long *max)
 227 {
 228         unsigned int boost_util, boost_max;
 229
 230         if (!sg_cpu->iowait_boost)
 231                 return;
 232
 233         if (sg_cpu->iowait_boost_pending) {
 234                 sg_cpu->iowait_boost_pending = false;
 235         } else {
 236                 sg_cpu->iowait_boost >>= 1;
 237                 if (sg_cpu->iowait_boost < sg_cpu->sg_policy->policy->min) {
 238                         sg_cpu->iowait_boost = 0;
 239                         return;
 240                 }
 241         }
 242
 243         boost_util = sg_cpu->iowait_boost;
 244         boost_max = sg_cpu->iowait_boost_max;
 245
 246         if (*util * boost_max < *max * boost_util) {
 247                 *util = boost_util;
 248                 *max = boost_max;
 249         }
 250 }
 251
 252 #ifdef CONFIG_NO_HZ_COMMON
 253 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
 254 {
 255         unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
 256         bool ret = idle_calls == sg_cpu->saved_idle_calls;
 257
 258         sg_cpu->saved_idle_calls = idle_calls;
 259         return ret;
 260 }
 261 #else
 262 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
 263 #endif /* CONFIG_NO_HZ_COMMON */
 264
 265 static void sugov_update_single(struct update_util_data *hook, u64 time,
 266                                 unsigned int flags)
 267 {
 268         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 269         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 270         struct cpufreq_policy *policy = sg_policy->policy;
 271         unsigned long util, max;
 272         unsigned int next_f;
 273         bool busy;
 274
 275         sugov_set_iowait_boost(sg_cpu, time);
 276         sg_cpu->last_update = time;
 277
 278         if (!sugov_should_update_freq(sg_policy, time))
 279                 return;
 280
 281         busy = sugov_cpu_is_busy(sg_cpu);
 282
 283         if (flags & SCHED_CPUFREQ_RT) {
 284                 next_f = policy->cpuinfo.max_freq;
 285         } else {
 286                 sugov_get_util(sg_cpu);
 287                 max = sg_cpu->max;
 288                 util = sugov_aggregate_util(sg_cpu);
 289                 sugov_iowait_boost(sg_cpu, &util, &max);
 290                 next_f = get_next_freq(sg_policy, util, max);
 291                 /*
 292                  * Do not reduce the frequency if the CPU has not been idle
 293                  * recently, as the reduction is likely to be premature then.
 294                  */
 295                 if (busy && next_f < sg_policy->next_freq) {
 296                         next_f = sg_policy->next_freq;
 297
 298                         /* Reset cached freq as next_freq has changed */
 299                         sg_policy->cached_raw_freq = 0;
 300                 }
 301         }
 302         sugov_update_commit(sg_policy, time, next_f);
 303 }
 304
 305 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
 306 {
 307         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 308         struct cpufreq_policy *policy = sg_policy->policy;
 309         unsigned long util = 0, max = 1;
 310         unsigned int j;
 311
 312         for_each_cpu(j, policy->cpus) {
 313                 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
 314                 unsigned long j_util, j_max;
 315                 s64 delta_ns;
 316
 317                 /*
 318                  * If the CFS CPU utilization was last updated before the
 319                  * previous frequency update and the time elapsed between the
 320                  * last update of the CPU utilization and the last frequency
 321                  * update is long enough, reset iowait_boost and util_cfs, as
 322                  * they are now probably stale. However, still consider the
 323                  * CPU contribution if it has some DEADLINE utilization
 324                  * (util_dl).
 325                  */
 326                 delta_ns = time - j_sg_cpu->last_update;
 327                 if (delta_ns > TICK_NSEC) {
 328                         j_sg_cpu->iowait_boost = 0;
 329                         j_sg_cpu->iowait_boost_pending = false;
 330                         j_sg_cpu->util_cfs = 0;
 331                         if (j_sg_cpu->util_dl == 0)
 332                                 continue;
 333                 }
 334                 if (j_sg_cpu->flags & SCHED_CPUFREQ_RT)
 335                         return policy->cpuinfo.max_freq;
 336
 337                 j_max = j_sg_cpu->max;
 338                 j_util = sugov_aggregate_util(j_sg_cpu);
 339                 if (j_util * max > j_max * util) {
 340                         util = j_util;
 341                         max = j_max;
 342                 }
 343
 344                 sugov_iowait_boost(j_sg_cpu, &util, &max);
 345         }
 346
 347         return get_next_freq(sg_policy, util, max);
 348 }
 349
 350 static void sugov_update_shared(struct update_util_data *hook, u64 time,
 351                                 unsigned int flags)
 352 {
 353         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 354         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 355         unsigned int next_f;
 356
 357         raw_spin_lock(&sg_policy->update_lock);
 358
 359         sugov_get_util(sg_cpu);
 360         sg_cpu->flags = flags;
 361
 362         sugov_set_iowait_boost(sg_cpu, time);
 363         sg_cpu->last_update = time;
 364
 365         if (sugov_should_update_freq(sg_policy, time)) {
 366                 if (flags & SCHED_CPUFREQ_RT)
 367                         next_f = sg_policy->policy->cpuinfo.max_freq;
 368                 else
 369                         next_f = sugov_next_freq_shared(sg_cpu, time);
 370
 371                 sugov_update_commit(sg_policy, time, next_f);
 372         }
 373
 374         raw_spin_unlock(&sg_policy->update_lock);
 375 }
 376
 377 static void sugov_work(struct kthread_work *work)
 378 {
 379         struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
 380
 381         mutex_lock(&sg_policy->work_lock);
 382         __cpufreq_driver_target(sg_policy->policy, sg_policy->next_freq,
 383                                 CPUFREQ_RELATION_L);
 384         mutex_unlock(&sg_policy->work_lock);
 385
 386         sg_policy->work_in_progress = false;
 387 }
 388
 389 static void sugov_irq_work(struct irq_work *irq_work)
 390 {
 391         struct sugov_policy *sg_policy;
 392
 393         sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
 394
 395         /*
 396          * For RT tasks, the schedutil governor shoots the frequency to maximum.
 397          * Special care must be taken to ensure that this kthread doesn't result
 398          * in the same behavior.
 399          *
 400          * This is (mostly) guaranteed by the work_in_progress flag. The flag is
 401          * updated only at the end of the sugov_work() function and before that
 402          * the schedutil governor rejects all other frequency scaling requests.
 403          *
 404          * There is a very rare case though, where the RT thread yields right
 405          * after the work_in_progress flag is cleared. The effects of that are
 406          * neglected for now.
 407          */
 408         kthread_queue_work(&sg_policy->worker, &sg_policy->work);
 409 }
 410
 411 /************************** sysfs interface ************************/
 412
 413 static struct sugov_tunables *global_tunables;
 414 static DEFINE_MUTEX(global_tunables_lock);
 415
 416 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
 417 {
 418         return container_of(attr_set, struct sugov_tunables, attr_set);
 419 }
 420
 421 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
 422 {
 423         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 424
 425         return sprintf(buf, "%u\n", tunables->rate_limit_us);
 426 }
 427
 428 static ssize_t rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf,
 429                                    size_t count)
 430 {
 431         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 432         struct sugov_policy *sg_policy;
 433         unsigned int rate_limit_us;
 434
 435         if (kstrtouint(buf, 10, &rate_limit_us))
 436                 return -EINVAL;
 437
 438         tunables->rate_limit_us = rate_limit_us;
 439
 440         list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
 441                 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
 442
 443         return count;
 444 }
 445
 446 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
 447
 448 static struct attribute *sugov_attributes[] = {
 449         &rate_limit_us.attr,
 450         NULL
 451 };
 452
 453 static struct kobj_type sugov_tunables_ktype = {
 454         .default_attrs = sugov_attributes,
 455         .sysfs_ops = &governor_sysfs_ops,
 456 };
 457
 458 /********************** cpufreq governor interface *********************/
 459
 460 static struct cpufreq_governor schedutil_gov;
 461
 462 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
 463 {
 464         struct sugov_policy *sg_policy;
 465
 466         sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
 467         if (!sg_policy)
 468                 return NULL;
 469
 470         sg_policy->policy = policy;
 471         raw_spin_lock_init(&sg_policy->update_lock);
 472         return sg_policy;
 473 }
 474
 475 static void sugov_policy_free(struct sugov_policy *sg_policy)
 476 {
 477         kfree(sg_policy);
 478 }
 479
 480 static int sugov_kthread_create(struct sugov_policy *sg_policy)
 481 {
 482         struct task_struct *thread;
 483         struct sched_attr attr = {
 484                 .size = sizeof(struct sched_attr),
 485                 .sched_policy = SCHED_DEADLINE,
 486                 .sched_flags = SCHED_FLAG_SUGOV,
 487                 .sched_nice = 0,
 488                 .sched_priority = 0,
 489                 /*
 490                  * Fake (unused) bandwidth; workaround to "fix"
 491                  * priority inheritance.
 492                  */
 493                 .sched_runtime  =  1000000,
 494                 .sched_deadline = 10000000,
 495                 .sched_period   = 10000000,
 496         };
 497         struct cpufreq_policy *policy = sg_policy->policy;
 498         int ret;
 499
 500         /* kthread only required for slow path */
 501         if (policy->fast_switch_enabled)
 502                 return 0;
 503
 504         kthread_init_work(&sg_policy->work, sugov_work);
 505         kthread_init_worker(&sg_policy->worker);
 506         thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
 507                                 "sugov:%d",
 508                                 cpumask_first(policy->related_cpus));
 509         if (IS_ERR(thread)) {
 510                 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
 511                 return PTR_ERR(thread);
 512         }
 513
 514         ret = sched_setattr_nocheck(thread, &attr);
 515         if (ret) {
 516                 kthread_stop(thread);
 517                 pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
 518                 return ret;
 519         }
 520
 521         sg_policy->thread = thread;
 522
 523         /* Kthread is bound to all CPUs by default */
 524         if (!policy->dvfs_possible_from_any_cpu)
 525                 kthread_bind_mask(thread, policy->related_cpus);
 526
 527         init_irq_work(&sg_policy->irq_work, sugov_irq_work);
 528         mutex_init(&sg_policy->work_lock);
 529
 530         wake_up_process(thread);
 531
 532         return 0;
 533 }
 534
 535 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
 536 {
 537         /* kthread only required for slow path */
 538         if (sg_policy->policy->fast_switch_enabled)
 539                 return;
 540
 541         kthread_flush_worker(&sg_policy->worker);
 542         kthread_stop(sg_policy->thread);
 543         mutex_destroy(&sg_policy->work_lock);
 544 }
 545
 546 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
 547 {
 548         struct sugov_tunables *tunables;
 549
 550         tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
 551         if (tunables) {
 552                 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
 553                 if (!have_governor_per_policy())
 554                         global_tunables = tunables;
 555         }
 556         return tunables;
 557 }
 558
 559 static void sugov_tunables_free(struct sugov_tunables *tunables)
 560 {
 561         if (!have_governor_per_policy())
 562                 global_tunables = NULL;
 563
 564         kfree(tunables);
 565 }
 566
 567 static int sugov_init(struct cpufreq_policy *policy)
 568 {
 569         struct sugov_policy *sg_policy;
 570         struct sugov_tunables *tunables;
 571         int ret = 0;
 572
 573         /* State should be equivalent to EXIT */
 574         if (policy->governor_data)
 575                 return -EBUSY;
 576
 577         cpufreq_enable_fast_switch(policy);
 578
 579         sg_policy = sugov_policy_alloc(policy);
 580         if (!sg_policy) {
 581                 ret = -ENOMEM;
 582                 goto disable_fast_switch;
 583         }
 584
 585         ret = sugov_kthread_create(sg_policy);
 586         if (ret)
 587                 goto free_sg_policy;
 588
 589         mutex_lock(&global_tunables_lock);
 590
 591         if (global_tunables) {
 592                 if (WARN_ON(have_governor_per_policy())) {
 593                         ret = -EINVAL;
 594                         goto stop_kthread;
 595                 }
 596                 policy->governor_data = sg_policy;
 597                 sg_policy->tunables = global_tunables;
 598
 599                 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
 600                 goto out;
 601         }
 602
 603         tunables = sugov_tunables_alloc(sg_policy);
 604         if (!tunables) {
 605                 ret = -ENOMEM;
 606                 goto stop_kthread;
 607         }
 608
 609         tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
 610
 611         policy->governor_data = sg_policy;
 612         sg_policy->tunables = tunables;
 613
 614         ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
 615                                    get_governor_parent_kobj(policy), "%s",
 616                                    schedutil_gov.name);
 617         if (ret)
 618                 goto fail;
 619
 620 out:
 621         mutex_unlock(&global_tunables_lock);
 622         return 0;
 623
 624 fail:
 625         policy->governor_data = NULL;
 626         sugov_tunables_free(tunables);
 627
 628 stop_kthread:
 629         sugov_kthread_stop(sg_policy);
 630
 631 free_sg_policy:
 632         mutex_unlock(&global_tunables_lock);
 633
 634         sugov_policy_free(sg_policy);
 635
 636 disable_fast_switch:
 637         cpufreq_disable_fast_switch(policy);
 638
 639         pr_err("initialization failed (error %d)\n", ret);
 640         return ret;
 641 }
 642
 643 static void sugov_exit(struct cpufreq_policy *policy)
 644 {
 645         struct sugov_policy *sg_policy = policy->governor_data;
 646         struct sugov_tunables *tunables = sg_policy->tunables;
 647         unsigned int count;
 648
 649         mutex_lock(&global_tunables_lock);
 650
 651         count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
 652         policy->governor_data = NULL;
 653         if (!count)
 654                 sugov_tunables_free(tunables);
 655
 656         mutex_unlock(&global_tunables_lock);
 657
 658         sugov_kthread_stop(sg_policy);
 659         sugov_policy_free(sg_policy);
 660         cpufreq_disable_fast_switch(policy);
 661 }
 662
 663 static int sugov_start(struct cpufreq_policy *policy)
 664 {
 665         struct sugov_policy *sg_policy = policy->governor_data;
 666         unsigned int cpu;
 667
 668         sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
 669         sg_policy->last_freq_update_time = 0;
 670         sg_policy->next_freq = UINT_MAX;
 671         sg_policy->work_in_progress = false;
 672         sg_policy->need_freq_update = false;
 673         sg_policy->cached_raw_freq = 0;
 674
 675         for_each_cpu(cpu, policy->cpus) {
 676                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 677
 678                 memset(sg_cpu, 0, sizeof(*sg_cpu));
 679                 sg_cpu->cpu = cpu;
 680                 sg_cpu->sg_policy = sg_policy;
 681                 sg_cpu->flags = 0;
 682                 sg_cpu->iowait_boost_max = policy->cpuinfo.max_freq;
 683         }
 684
 685         for_each_cpu(cpu, policy->cpus) {
 686                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 687
 688                 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
 689                                              policy_is_shared(policy) ?
 690                                                         sugov_update_shared :
 691                                                         sugov_update_single);
 692         }
 693         return 0;
 694 }
 695
 696 static void sugov_stop(struct cpufreq_policy *policy)
 697 {
 698         struct sugov_policy *sg_policy = policy->governor_data;
 699         unsigned int cpu;
 700
 701         for_each_cpu(cpu, policy->cpus)
 702                 cpufreq_remove_update_util_hook(cpu);
 703
 704         synchronize_sched();
 705
 706         if (!policy->fast_switch_enabled) {
 707                 irq_work_sync(&sg_policy->irq_work);
 708                 kthread_cancel_work_sync(&sg_policy->work);
 709         }
 710 }
 711
 712 static void sugov_limits(struct cpufreq_policy *policy)
 713 {
 714         struct sugov_policy *sg_policy = policy->governor_data;
 715
 716         if (!policy->fast_switch_enabled) {
 717                 mutex_lock(&sg_policy->work_lock);
 718                 cpufreq_policy_apply_limits(policy);
 719                 mutex_unlock(&sg_policy->work_lock);
 720         }
 721
 722         sg_policy->need_freq_update = true;
 723 }
 724
 725 static struct cpufreq_governor schedutil_gov = {
 726         .name = "schedutil",
 727         .owner = THIS_MODULE,
 728         .dynamic_switching = true,
 729         .init = sugov_init,
 730         .exit = sugov_exit,
 731         .start = sugov_start,
 732         .stop = sugov_stop,
 733         .limits = sugov_limits,
 734 };
 735
 736 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
 737 struct cpufreq_governor *cpufreq_default_governor(void)
 738 {
 739         return &schedutil_gov;
 740 }
 741 #endif
 742
 743 static int __init sugov_register(void)
 744 {
 745         return cpufreq_register_governor(&schedutil_gov);
 746 }
 747 fs_initcall(sugov_register);