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 struct sugov_tunables {
  23         struct gov_attr_set attr_set;
  24         unsigned int rate_limit_us;
  25 };
  26
  27 struct sugov_policy {
  28         struct cpufreq_policy *policy;
  29
  30         struct sugov_tunables *tunables;
  31         struct list_head tunables_hook;
  32
  33         raw_spinlock_t update_lock;  /* For shared policies */
  34         u64 last_freq_update_time;
  35         s64 freq_update_delay_ns;
  36         unsigned int next_freq;
  37         unsigned int cached_raw_freq;
  38
  39         /* The next fields are only needed if fast switch cannot be used. */
  40         struct irq_work irq_work;
  41         struct kthread_work work;
  42         struct mutex work_lock;
  43         struct kthread_worker worker;
  44         struct task_struct *thread;
  45         bool work_in_progress;
  46
  47         bool need_freq_update;
  48 };
  49
  50 struct sugov_cpu {
  51         struct update_util_data update_util;
  52         struct sugov_policy *sg_policy;
  53         unsigned int cpu;
  54
  55         bool iowait_boost_pending;
  56         unsigned int iowait_boost;
  57         unsigned int iowait_boost_max;
  58         u64 last_update;
  59
  60         /* The fields below are only needed when sharing a policy. */
  61         unsigned long util_cfs;
  62         unsigned long util_dl;
  63         unsigned long max;
  64         unsigned int flags;
  65
  66         /* The field below is for single-CPU policies only. */
  67 #ifdef CONFIG_NO_HZ_COMMON
  68         unsigned long saved_idle_calls;
  69 #endif
  70 };
  71
  72 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
  73
  74 /************************ Governor internals ***********************/
  75
  76 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
  77 {
  78         s64 delta_ns;
  79
  80         /*
  81          * Since cpufreq_update_util() is called with rq->lock held for
  82          * the @target_cpu, our per-cpu data is fully serialized.
  83          *
  84          * However, drivers cannot in general deal with cross-cpu
  85          * requests, so while get_next_freq() will work, our
  86          * sugov_update_commit() call may not for the fast switching platforms.
  87          *
  88          * Hence stop here for remote requests if they aren't supported
  89          * by the hardware, as calculating the frequency is pointless if
  90          * we cannot in fact act on it.
  91          *
  92          * For the slow switching platforms, the kthread is always scheduled on
  93          * the right set of CPUs and any CPU can find the next frequency and
  94          * schedule the kthread.
  95          */
  96         if (sg_policy->policy->fast_switch_enabled &&
  97             !cpufreq_can_do_remote_dvfs(sg_policy->policy))
  98                 return false;
  99
 100         if (sg_policy->work_in_progress)
 101                 return false;
 102
 103         if (unlikely(sg_policy->need_freq_update)) {
 104                 sg_policy->need_freq_update = false;
 105                 /*
 106                  * This happens when limits change, so forget the previous
 107                  * next_freq value and force an update.
 108                  */
 109                 sg_policy->next_freq = UINT_MAX;
 110                 return true;
 111         }
 112
 113         delta_ns = time - sg_policy->last_freq_update_time;
 114         return delta_ns >= sg_policy->freq_update_delay_ns;
 115 }
 116
 117 static void sugov_update_commit(struct sugov_policy *sg_policy, u64 time,
 118                                 unsigned int next_freq)
 119 {
 120         struct cpufreq_policy *policy = sg_policy->policy;
 121
 122         if (sg_policy->next_freq == next_freq)
 123                 return;
 124
 125         sg_policy->next_freq = next_freq;
 126         sg_policy->last_freq_update_time = time;
 127
 128         if (policy->fast_switch_enabled) {
 129                 next_freq = cpufreq_driver_fast_switch(policy, next_freq);
 130                 if (!next_freq)
 131                         return;
 132
 133                 policy->cur = next_freq;
 134                 trace_cpu_frequency(next_freq, smp_processor_id());
 135         } else {
 136                 sg_policy->work_in_progress = true;
 137                 irq_work_queue(&sg_policy->irq_work);
 138         }
 139 }
 140
 141 /**
 142  * get_next_freq - Compute a new frequency for a given cpufreq policy.
 143  * @sg_policy: schedutil policy object to compute the new frequency for.
 144  * @util: Current CPU utilization.
 145  * @max: CPU capacity.
 146  *
 147  * If the utilization is frequency-invariant, choose the new frequency to be
 148  * proportional to it, that is
 149  *
 150  * next_freq = C * max_freq * util / max
 151  *
 152  * Otherwise, approximate the would-be frequency-invariant utilization by
 153  * util_raw * (curr_freq / max_freq) which leads to
 154  *
 155  * next_freq = C * curr_freq * util_raw / max
 156  *
 157  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
 158  *
 159  * The lowest driver-supported frequency which is equal or greater than the raw
 160  * next_freq (as calculated above) is returned, subject to policy min/max and
 161  * cpufreq driver limitations.
 162  */
 163 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
 164                                   unsigned long util, unsigned long max)
 165 {
 166         struct cpufreq_policy *policy = sg_policy->policy;
 167         unsigned int freq = arch_scale_freq_invariant() ?
 168                                 policy->cpuinfo.max_freq : policy->cur;
 169
 170         freq = (freq + (freq >> 2)) * util / max;
 171
 172         if (freq == sg_policy->cached_raw_freq && sg_policy->next_freq != UINT_MAX)
 173                 return sg_policy->next_freq;
 174         sg_policy->cached_raw_freq = freq;
 175         return cpufreq_driver_resolve_freq(policy, freq);
 176 }
 177
 178 static void sugov_get_util(struct sugov_cpu *sg_cpu)
 179 {
 180         struct rq *rq = cpu_rq(sg_cpu->cpu);
 181
 182         sg_cpu->max = arch_scale_cpu_capacity(NULL, sg_cpu->cpu);
 183         sg_cpu->util_cfs = cpu_util_cfs(rq);
 184         sg_cpu->util_dl  = cpu_util_dl(rq);
 185 }
 186
 187 static unsigned long sugov_aggregate_util(struct sugov_cpu *sg_cpu)
 188 {
 189         /*
 190          * Ideally we would like to set util_dl as min/guaranteed freq and
 191          * util_cfs + util_dl as requested freq. However, cpufreq is not yet
 192          * ready for such an interface. So, we only do the latter for now.
 193          */
 194         return min(sg_cpu->util_cfs + sg_cpu->util_dl, sg_cpu->max);
 195 }
 196
 197 static void sugov_set_iowait_boost(struct sugov_cpu *sg_cpu, u64 time)
 198 {
 199         if (sg_cpu->flags & SCHED_CPUFREQ_IOWAIT) {
 200                 if (sg_cpu->iowait_boost_pending)
 201                         return;
 202
 203                 sg_cpu->iowait_boost_pending = true;
 204
 205                 if (sg_cpu->iowait_boost) {
 206                         sg_cpu->iowait_boost <<= 1;
 207                         if (sg_cpu->iowait_boost > sg_cpu->iowait_boost_max)
 208                                 sg_cpu->iowait_boost = sg_cpu->iowait_boost_max;
 209                 } else {
 210                         sg_cpu->iowait_boost = sg_cpu->sg_policy->policy->min;
 211                 }
 212         } else if (sg_cpu->iowait_boost) {
 213                 s64 delta_ns = time - sg_cpu->last_update;
 214
 215                 /* Clear iowait_boost if the CPU apprears to have been idle. */
 216                 if (delta_ns > TICK_NSEC) {
 217                         sg_cpu->iowait_boost = 0;
 218                         sg_cpu->iowait_boost_pending = false;
 219                 }
 220         }
 221 }
 222
 223 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, unsigned long *util,
 224                                unsigned long *max)
 225 {
 226         unsigned int boost_util, boost_max;
 227
 228         if (!sg_cpu->iowait_boost)
 229                 return;
 230
 231         if (sg_cpu->iowait_boost_pending) {
 232                 sg_cpu->iowait_boost_pending = false;
 233         } else {
 234                 sg_cpu->iowait_boost >>= 1;
 235                 if (sg_cpu->iowait_boost < sg_cpu->sg_policy->policy->min) {
 236                         sg_cpu->iowait_boost = 0;
 237                         return;
 238                 }
 239         }
 240
 241         boost_util = sg_cpu->iowait_boost;
 242         boost_max = sg_cpu->iowait_boost_max;
 243
 244         if (*util * boost_max < *max * boost_util) {
 245                 *util = boost_util;
 246                 *max = boost_max;
 247         }
 248 }
 249
 250 #ifdef CONFIG_NO_HZ_COMMON
 251 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
 252 {
 253         unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
 254         bool ret = idle_calls == sg_cpu->saved_idle_calls;
 255
 256         sg_cpu->saved_idle_calls = idle_calls;
 257         return ret;
 258 }
 259 #else
 260 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
 261 #endif /* CONFIG_NO_HZ_COMMON */
 262
 263 static void sugov_update_single(struct update_util_data *hook, u64 time,
 264                                 unsigned int flags)
 265 {
 266         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 267         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 268         struct cpufreq_policy *policy = sg_policy->policy;
 269         unsigned long util, max;
 270         unsigned int next_f;
 271         bool busy;
 272
 273         sugov_set_iowait_boost(sg_cpu, time);
 274         sg_cpu->last_update = time;
 275
 276         if (!sugov_should_update_freq(sg_policy, time))
 277                 return;
 278
 279         busy = sugov_cpu_is_busy(sg_cpu);
 280
 281         if (flags & SCHED_CPUFREQ_RT) {
 282                 next_f = policy->cpuinfo.max_freq;
 283         } else {
 284                 sugov_get_util(sg_cpu);
 285                 max = sg_cpu->max;
 286                 util = sugov_aggregate_util(sg_cpu);
 287                 sugov_iowait_boost(sg_cpu, &util, &max);
 288                 next_f = get_next_freq(sg_policy, util, max);
 289                 /*
 290                  * Do not reduce the frequency if the CPU has not been idle
 291                  * recently, as the reduction is likely to be premature then.
 292                  */
 293                 if (busy && next_f < sg_policy->next_freq) {
 294                         next_f = sg_policy->next_freq;
 295
 296                         /* Reset cached freq as next_freq has changed */
 297                         sg_policy->cached_raw_freq = 0;
 298                 }
 299         }
 300         sugov_update_commit(sg_policy, time, next_f);
 301 }
 302
 303 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
 304 {
 305         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 306         struct cpufreq_policy *policy = sg_policy->policy;
 307         unsigned long util = 0, max = 1;
 308         unsigned int j;
 309
 310         for_each_cpu(j, policy->cpus) {
 311                 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
 312                 unsigned long j_util, j_max;
 313                 s64 delta_ns;
 314
 315                 /*
 316                  * If the CFS CPU utilization was last updated before the
 317                  * previous frequency update and the time elapsed between the
 318                  * last update of the CPU utilization and the last frequency
 319                  * update is long enough, reset iowait_boost and util_cfs, as
 320                  * they are now probably stale. However, still consider the
 321                  * CPU contribution if it has some DEADLINE utilization
 322                  * (util_dl).
 323                  */
 324                 delta_ns = time - j_sg_cpu->last_update;
 325                 if (delta_ns > TICK_NSEC) {
 326                         j_sg_cpu->iowait_boost = 0;
 327                         j_sg_cpu->iowait_boost_pending = false;
 328                         j_sg_cpu->util_cfs = 0;
 329                         if (j_sg_cpu->util_dl == 0)
 330                                 continue;
 331                 }
 332                 if (j_sg_cpu->flags & SCHED_CPUFREQ_RT)
 333                         return policy->cpuinfo.max_freq;
 334
 335                 j_max = j_sg_cpu->max;
 336                 j_util = sugov_aggregate_util(j_sg_cpu);
 337                 if (j_util * max > j_max * util) {
 338                         util = j_util;
 339                         max = j_max;
 340                 }
 341
 342                 sugov_iowait_boost(j_sg_cpu, &util, &max);
 343         }
 344
 345         return get_next_freq(sg_policy, util, max);
 346 }
 347
 348 static void sugov_update_shared(struct update_util_data *hook, u64 time,
 349                                 unsigned int flags)
 350 {
 351         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 352         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 353         unsigned int next_f;
 354
 355         raw_spin_lock(&sg_policy->update_lock);
 356
 357         sugov_get_util(sg_cpu);
 358         sg_cpu->flags = flags;
 359
 360         sugov_set_iowait_boost(sg_cpu, time);
 361         sg_cpu->last_update = time;
 362
 363         if (sugov_should_update_freq(sg_policy, time)) {
 364                 if (flags & SCHED_CPUFREQ_RT)
 365                         next_f = sg_policy->policy->cpuinfo.max_freq;
 366                 else
 367                         next_f = sugov_next_freq_shared(sg_cpu, time);
 368
 369                 sugov_update_commit(sg_policy, time, next_f);
 370         }
 371
 372         raw_spin_unlock(&sg_policy->update_lock);
 373 }
 374
 375 static void sugov_work(struct kthread_work *work)
 376 {
 377         struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
 378
 379         mutex_lock(&sg_policy->work_lock);
 380         __cpufreq_driver_target(sg_policy->policy, sg_policy->next_freq,
 381                                 CPUFREQ_RELATION_L);
 382         mutex_unlock(&sg_policy->work_lock);
 383
 384         sg_policy->work_in_progress = false;
 385 }
 386
 387 static void sugov_irq_work(struct irq_work *irq_work)
 388 {
 389         struct sugov_policy *sg_policy;
 390
 391         sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
 392
 393         /*
 394          * For RT tasks, the schedutil governor shoots the frequency to maximum.
 395          * Special care must be taken to ensure that this kthread doesn't result
 396          * in the same behavior.
 397          *
 398          * This is (mostly) guaranteed by the work_in_progress flag. The flag is
 399          * updated only at the end of the sugov_work() function and before that
 400          * the schedutil governor rejects all other frequency scaling requests.
 401          *
 402          * There is a very rare case though, where the RT thread yields right
 403          * after the work_in_progress flag is cleared. The effects of that are
 404          * neglected for now.
 405          */
 406         kthread_queue_work(&sg_policy->worker, &sg_policy->work);
 407 }
 408
 409 /************************** sysfs interface ************************/
 410
 411 static struct sugov_tunables *global_tunables;
 412 static DEFINE_MUTEX(global_tunables_lock);
 413
 414 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
 415 {
 416         return container_of(attr_set, struct sugov_tunables, attr_set);
 417 }
 418
 419 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
 420 {
 421         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 422
 423         return sprintf(buf, "%u\n", tunables->rate_limit_us);
 424 }
 425
 426 static ssize_t rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf,
 427                                    size_t count)
 428 {
 429         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 430         struct sugov_policy *sg_policy;
 431         unsigned int rate_limit_us;
 432
 433         if (kstrtouint(buf, 10, &rate_limit_us))
 434                 return -EINVAL;
 435
 436         tunables->rate_limit_us = rate_limit_us;
 437
 438         list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
 439                 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
 440
 441         return count;
 442 }
 443
 444 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
 445
 446 static struct attribute *sugov_attributes[] = {
 447         &rate_limit_us.attr,
 448         NULL
 449 };
 450
 451 static struct kobj_type sugov_tunables_ktype = {
 452         .default_attrs = sugov_attributes,
 453         .sysfs_ops = &governor_sysfs_ops,
 454 };
 455
 456 /********************** cpufreq governor interface *********************/
 457
 458 static struct cpufreq_governor schedutil_gov;
 459
 460 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
 461 {
 462         struct sugov_policy *sg_policy;
 463
 464         sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
 465         if (!sg_policy)
 466                 return NULL;
 467
 468         sg_policy->policy = policy;
 469         raw_spin_lock_init(&sg_policy->update_lock);
 470         return sg_policy;
 471 }
 472
 473 static void sugov_policy_free(struct sugov_policy *sg_policy)
 474 {
 475         kfree(sg_policy);
 476 }
 477
 478 static int sugov_kthread_create(struct sugov_policy *sg_policy)
 479 {
 480         struct task_struct *thread;
 481         struct sched_attr attr = {
 482                 .size = sizeof(struct sched_attr),
 483                 .sched_policy = SCHED_DEADLINE,
 484                 .sched_flags = SCHED_FLAG_SUGOV,
 485                 .sched_nice = 0,
 486                 .sched_priority = 0,
 487                 /*
 488                  * Fake (unused) bandwidth; workaround to "fix"
 489                  * priority inheritance.
 490                  */
 491                 .sched_runtime  =  1000000,
 492                 .sched_deadline = 10000000,
 493                 .sched_period   = 10000000,
 494         };
 495         struct cpufreq_policy *policy = sg_policy->policy;
 496         int ret;
 497
 498         /* kthread only required for slow path */
 499         if (policy->fast_switch_enabled)
 500                 return 0;
 501
 502         kthread_init_work(&sg_policy->work, sugov_work);
 503         kthread_init_worker(&sg_policy->worker);
 504         thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
 505                                 "sugov:%d",
 506                                 cpumask_first(policy->related_cpus));
 507         if (IS_ERR(thread)) {
 508                 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
 509                 return PTR_ERR(thread);
 510         }
 511
 512         ret = sched_setattr_nocheck(thread, &attr);
 513         if (ret) {
 514                 kthread_stop(thread);
 515                 pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
 516                 return ret;
 517         }
 518
 519         sg_policy->thread = thread;
 520
 521         /* Kthread is bound to all CPUs by default */
 522         if (!policy->dvfs_possible_from_any_cpu)
 523                 kthread_bind_mask(thread, policy->related_cpus);
 524
 525         init_irq_work(&sg_policy->irq_work, sugov_irq_work);
 526         mutex_init(&sg_policy->work_lock);
 527
 528         wake_up_process(thread);
 529
 530         return 0;
 531 }
 532
 533 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
 534 {
 535         /* kthread only required for slow path */
 536         if (sg_policy->policy->fast_switch_enabled)
 537                 return;
 538
 539         kthread_flush_worker(&sg_policy->worker);
 540         kthread_stop(sg_policy->thread);
 541         mutex_destroy(&sg_policy->work_lock);
 542 }
 543
 544 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
 545 {
 546         struct sugov_tunables *tunables;
 547
 548         tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
 549         if (tunables) {
 550                 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
 551                 if (!have_governor_per_policy())
 552                         global_tunables = tunables;
 553         }
 554         return tunables;
 555 }
 556
 557 static void sugov_tunables_free(struct sugov_tunables *tunables)
 558 {
 559         if (!have_governor_per_policy())
 560                 global_tunables = NULL;
 561
 562         kfree(tunables);
 563 }
 564
 565 static int sugov_init(struct cpufreq_policy *policy)
 566 {
 567         struct sugov_policy *sg_policy;
 568         struct sugov_tunables *tunables;
 569         int ret = 0;
 570
 571         /* State should be equivalent to EXIT */
 572         if (policy->governor_data)
 573                 return -EBUSY;
 574
 575         cpufreq_enable_fast_switch(policy);
 576
 577         sg_policy = sugov_policy_alloc(policy);
 578         if (!sg_policy) {
 579                 ret = -ENOMEM;
 580                 goto disable_fast_switch;
 581         }
 582
 583         ret = sugov_kthread_create(sg_policy);
 584         if (ret)
 585                 goto free_sg_policy;
 586
 587         mutex_lock(&global_tunables_lock);
 588
 589         if (global_tunables) {
 590                 if (WARN_ON(have_governor_per_policy())) {
 591                         ret = -EINVAL;
 592                         goto stop_kthread;
 593                 }
 594                 policy->governor_data = sg_policy;
 595                 sg_policy->tunables = global_tunables;
 596
 597                 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
 598                 goto out;
 599         }
 600
 601         tunables = sugov_tunables_alloc(sg_policy);
 602         if (!tunables) {
 603                 ret = -ENOMEM;
 604                 goto stop_kthread;
 605         }
 606
 607         tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
 608
 609         policy->governor_data = sg_policy;
 610         sg_policy->tunables = tunables;
 611
 612         ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
 613                                    get_governor_parent_kobj(policy), "%s",
 614                                    schedutil_gov.name);
 615         if (ret)
 616                 goto fail;
 617
 618 out:
 619         mutex_unlock(&global_tunables_lock);
 620         return 0;
 621
 622 fail:
 623         policy->governor_data = NULL;
 624         sugov_tunables_free(tunables);
 625
 626 stop_kthread:
 627         sugov_kthread_stop(sg_policy);
 628
 629 free_sg_policy:
 630         mutex_unlock(&global_tunables_lock);
 631
 632         sugov_policy_free(sg_policy);
 633
 634 disable_fast_switch:
 635         cpufreq_disable_fast_switch(policy);
 636
 637         pr_err("initialization failed (error %d)\n", ret);
 638         return ret;
 639 }
 640
 641 static void sugov_exit(struct cpufreq_policy *policy)
 642 {
 643         struct sugov_policy *sg_policy = policy->governor_data;
 644         struct sugov_tunables *tunables = sg_policy->tunables;
 645         unsigned int count;
 646
 647         mutex_lock(&global_tunables_lock);
 648
 649         count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
 650         policy->governor_data = NULL;
 651         if (!count)
 652                 sugov_tunables_free(tunables);
 653
 654         mutex_unlock(&global_tunables_lock);
 655
 656         sugov_kthread_stop(sg_policy);
 657         sugov_policy_free(sg_policy);
 658         cpufreq_disable_fast_switch(policy);
 659 }
 660
 661 static int sugov_start(struct cpufreq_policy *policy)
 662 {
 663         struct sugov_policy *sg_policy = policy->governor_data;
 664         unsigned int cpu;
 665
 666         sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
 667         sg_policy->last_freq_update_time = 0;
 668         sg_policy->next_freq = UINT_MAX;
 669         sg_policy->work_in_progress = false;
 670         sg_policy->need_freq_update = false;
 671         sg_policy->cached_raw_freq = 0;
 672
 673         for_each_cpu(cpu, policy->cpus) {
 674                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 675
 676                 memset(sg_cpu, 0, sizeof(*sg_cpu));
 677                 sg_cpu->cpu = cpu;
 678                 sg_cpu->sg_policy = sg_policy;
 679                 sg_cpu->flags = 0;
 680                 sg_cpu->iowait_boost_max = policy->cpuinfo.max_freq;
 681         }
 682
 683         for_each_cpu(cpu, policy->cpus) {
 684                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 685
 686                 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
 687                                              policy_is_shared(policy) ?
 688                                                         sugov_update_shared :
 689                                                         sugov_update_single);
 690         }
 691         return 0;
 692 }
 693
 694 static void sugov_stop(struct cpufreq_policy *policy)
 695 {
 696         struct sugov_policy *sg_policy = policy->governor_data;
 697         unsigned int cpu;
 698
 699         for_each_cpu(cpu, policy->cpus)
 700                 cpufreq_remove_update_util_hook(cpu);
 701
 702         synchronize_sched();
 703
 704         if (!policy->fast_switch_enabled) {
 705                 irq_work_sync(&sg_policy->irq_work);
 706                 kthread_cancel_work_sync(&sg_policy->work);
 707         }
 708 }
 709
 710 static void sugov_limits(struct cpufreq_policy *policy)
 711 {
 712         struct sugov_policy *sg_policy = policy->governor_data;
 713
 714         if (!policy->fast_switch_enabled) {
 715                 mutex_lock(&sg_policy->work_lock);
 716                 cpufreq_policy_apply_limits(policy);
 717                 mutex_unlock(&sg_policy->work_lock);
 718         }
 719
 720         sg_policy->need_freq_update = true;
 721 }
 722
 723 static struct cpufreq_governor schedutil_gov = {
 724         .name = "schedutil",
 725         .owner = THIS_MODULE,
 726         .dynamic_switching = true,
 727         .init = sugov_init,
 728         .exit = sugov_exit,
 729         .start = sugov_start,
 730         .stop = sugov_stop,
 731         .limits = sugov_limits,
 732 };
 733
 734 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
 735 struct cpufreq_governor *cpufreq_default_governor(void)
 736 {
 737         return &schedutil_gov;
 738 }
 739 #endif
 740
 741 static int __init sugov_register(void)
 742 {
 743         return cpufreq_register_governor(&schedutil_gov);
 744 }
 745 fs_initcall(sugov_register);