kernel/sched/cpufreq_schedutil.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * CPUFreq governor based on scheduler-provided CPU utilization data.
   4  *
   5  * Copyright (C) 2016, Intel Corporation
   6  * Author: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
   7  */
   8
   9 #define IOWAIT_BOOST_MIN        (SCHED_CAPACITY_SCALE / 8)
  10
  11 struct sugov_tunables {
  12         struct gov_attr_set     attr_set;
  13         unsigned int            rate_limit_us;
  14 };
  15
  16 struct sugov_policy {
  17         struct cpufreq_policy   *policy;
  18
  19         struct sugov_tunables   *tunables;
  20         struct list_head        tunables_hook;
  21
  22         raw_spinlock_t          update_lock;
  23         u64                     last_freq_update_time;
  24         s64                     freq_update_delay_ns;
  25         unsigned int            next_freq;
  26         unsigned int            cached_raw_freq;
  27
  28         /* The next fields are only needed if fast switch cannot be used: */
  29         struct                  irq_work irq_work;
  30         struct                  kthread_work work;
  31         struct                  mutex work_lock;
  32         struct                  kthread_worker worker;
  33         struct task_struct      *thread;
  34         bool                    work_in_progress;
  35
  36         bool                    limits_changed;
  37         bool                    need_freq_update;
  38 };
  39
  40 struct sugov_cpu {
  41         struct update_util_data update_util;
  42         struct sugov_policy     *sg_policy;
  43         unsigned int            cpu;
  44
  45         bool                    iowait_boost_pending;
  46         unsigned int            iowait_boost;
  47         u64                     last_update;
  48
  49         unsigned long           util;
  50         unsigned long           bw_dl;
  51
  52         /* The field below is for single-CPU policies only: */
  53 #ifdef CONFIG_NO_HZ_COMMON
  54         unsigned long           saved_idle_calls;
  55 #endif
  56 };
  57
  58 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
  59
  60 /************************ Governor internals ***********************/
  61
  62 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
  63 {
  64         s64 delta_ns;
  65
  66         /*
  67          * Since cpufreq_update_util() is called with rq->lock held for
  68          * the @target_cpu, our per-CPU data is fully serialized.
  69          *
  70          * However, drivers cannot in general deal with cross-CPU
  71          * requests, so while get_next_freq() will work, our
  72          * sugov_update_commit() call may not for the fast switching platforms.
  73          *
  74          * Hence stop here for remote requests if they aren't supported
  75          * by the hardware, as calculating the frequency is pointless if
  76          * we cannot in fact act on it.
  77          *
  78          * This is needed on the slow switching platforms too to prevent CPUs
  79          * going offline from leaving stale IRQ work items behind.
  80          */
  81         if (!cpufreq_this_cpu_can_update(sg_policy->policy))
  82                 return false;
  83
  84         if (unlikely(sg_policy->limits_changed)) {
  85                 sg_policy->limits_changed = false;
  86                 sg_policy->need_freq_update = true;
  87                 return true;
  88         }
  89
  90         delta_ns = time - sg_policy->last_freq_update_time;
  91
  92         return delta_ns >= sg_policy->freq_update_delay_ns;
  93 }
  94
  95 static bool sugov_update_next_freq(struct sugov_policy *sg_policy, u64 time,
  96                                    unsigned int next_freq)
  97 {
  98         if (sg_policy->need_freq_update)
  99                 sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
 100         else if (sg_policy->next_freq == next_freq)
 101                 return false;
 102
 103         sg_policy->next_freq = next_freq;
 104         sg_policy->last_freq_update_time = time;
 105
 106         return true;
 107 }
 108
 109 static void sugov_deferred_update(struct sugov_policy *sg_policy)
 110 {
 111         if (!sg_policy->work_in_progress) {
 112                 sg_policy->work_in_progress = true;
 113                 irq_work_queue(&sg_policy->irq_work);
 114         }
 115 }
 116
 117 /**
 118  * get_next_freq - Compute a new frequency for a given cpufreq policy.
 119  * @sg_policy: schedutil policy object to compute the new frequency for.
 120  * @util: Current CPU utilization.
 121  * @max: CPU capacity.
 122  *
 123  * If the utilization is frequency-invariant, choose the new frequency to be
 124  * proportional to it, that is
 125  *
 126  * next_freq = C * max_freq * util / max
 127  *
 128  * Otherwise, approximate the would-be frequency-invariant utilization by
 129  * util_raw * (curr_freq / max_freq) which leads to
 130  *
 131  * next_freq = C * curr_freq * util_raw / max
 132  *
 133  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
 134  *
 135  * The lowest driver-supported frequency which is equal or greater than the raw
 136  * next_freq (as calculated above) is returned, subject to policy min/max and
 137  * cpufreq driver limitations.
 138  */
 139 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
 140                                   unsigned long util, unsigned long max)
 141 {
 142         struct cpufreq_policy *policy = sg_policy->policy;
 143         unsigned int freq = arch_scale_freq_invariant() ?
 144                                 policy->cpuinfo.max_freq : policy->cur;
 145
 146         util = map_util_perf(util);
 147         freq = map_util_freq(util, freq, max);
 148
 149         if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
 150                 return sg_policy->next_freq;
 151
 152         sg_policy->cached_raw_freq = freq;
 153         return cpufreq_driver_resolve_freq(policy, freq);
 154 }
 155
 156 static void sugov_get_util(struct sugov_cpu *sg_cpu)
 157 {
 158         unsigned long util = cpu_util_cfs_boost(sg_cpu->cpu);
 159         struct rq *rq = cpu_rq(sg_cpu->cpu);
 160
 161         sg_cpu->bw_dl = cpu_bw_dl(rq);
 162         sg_cpu->util = effective_cpu_util(sg_cpu->cpu, util,
 163                                           FREQUENCY_UTIL, NULL);
 164 }
 165
 166 /**
 167  * sugov_iowait_reset() - Reset the IO boost status of a CPU.
 168  * @sg_cpu: the sugov data for the CPU to boost
 169  * @time: the update time from the caller
 170  * @set_iowait_boost: true if an IO boost has been requested
 171  *
 172  * The IO wait boost of a task is disabled after a tick since the last update
 173  * of a CPU. If a new IO wait boost is requested after more then a tick, then
 174  * we enable the boost starting from IOWAIT_BOOST_MIN, which improves energy
 175  * efficiency by ignoring sporadic wakeups from IO.
 176  */
 177 static bool sugov_iowait_reset(struct sugov_cpu *sg_cpu, u64 time,
 178                                bool set_iowait_boost)
 179 {
 180         s64 delta_ns = time - sg_cpu->last_update;
 181
 182         /* Reset boost only if a tick has elapsed since last request */
 183         if (delta_ns <= TICK_NSEC)
 184                 return false;
 185
 186         sg_cpu->iowait_boost = set_iowait_boost ? IOWAIT_BOOST_MIN : 0;
 187         sg_cpu->iowait_boost_pending = set_iowait_boost;
 188
 189         return true;
 190 }
 191
 192 /**
 193  * sugov_iowait_boost() - Updates the IO boost status of a CPU.
 194  * @sg_cpu: the sugov data for the CPU to boost
 195  * @time: the update time from the caller
 196  * @flags: SCHED_CPUFREQ_IOWAIT if the task is waking up after an IO wait
 197  *
 198  * Each time a task wakes up after an IO operation, the CPU utilization can be
 199  * boosted to a certain utilization which doubles at each "frequent and
 200  * successive" wakeup from IO, ranging from IOWAIT_BOOST_MIN to the utilization
 201  * of the maximum OPP.
 202  *
 203  * To keep doubling, an IO boost has to be requested at least once per tick,
 204  * otherwise we restart from the utilization of the minimum OPP.
 205  */
 206 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
 207                                unsigned int flags)
 208 {
 209         bool set_iowait_boost = flags & SCHED_CPUFREQ_IOWAIT;
 210
 211         /* Reset boost if the CPU appears to have been idle enough */
 212         if (sg_cpu->iowait_boost &&
 213             sugov_iowait_reset(sg_cpu, time, set_iowait_boost))
 214                 return;
 215
 216         /* Boost only tasks waking up after IO */
 217         if (!set_iowait_boost)
 218                 return;
 219
 220         /* Ensure boost doubles only one time at each request */
 221         if (sg_cpu->iowait_boost_pending)
 222                 return;
 223         sg_cpu->iowait_boost_pending = true;
 224
 225         /* Double the boost at each request */
 226         if (sg_cpu->iowait_boost) {
 227                 sg_cpu->iowait_boost =
 228                         min_t(unsigned int, sg_cpu->iowait_boost << 1, SCHED_CAPACITY_SCALE);
 229                 return;
 230         }
 231
 232         /* First wakeup after IO: start with minimum boost */
 233         sg_cpu->iowait_boost = IOWAIT_BOOST_MIN;
 234 }
 235
 236 /**
 237  * sugov_iowait_apply() - Apply the IO boost to a CPU.
 238  * @sg_cpu: the sugov data for the cpu to boost
 239  * @time: the update time from the caller
 240  * @max_cap: the max CPU capacity
 241  *
 242  * A CPU running a task which woken up after an IO operation can have its
 243  * utilization boosted to speed up the completion of those IO operations.
 244  * The IO boost value is increased each time a task wakes up from IO, in
 245  * sugov_iowait_apply(), and it's instead decreased by this function,
 246  * each time an increase has not been requested (!iowait_boost_pending).
 247  *
 248  * A CPU which also appears to have been idle for at least one tick has also
 249  * its IO boost utilization reset.
 250  *
 251  * This mechanism is designed to boost high frequently IO waiting tasks, while
 252  * being more conservative on tasks which does sporadic IO operations.
 253  */
 254 static void sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
 255                                unsigned long max_cap)
 256 {
 257         unsigned long boost;
 258
 259         /* No boost currently required */
 260         if (!sg_cpu->iowait_boost)
 261                 return;
 262
 263         /* Reset boost if the CPU appears to have been idle enough */
 264         if (sugov_iowait_reset(sg_cpu, time, false))
 265                 return;
 266
 267         if (!sg_cpu->iowait_boost_pending) {
 268                 /*
 269                  * No boost pending; reduce the boost value.
 270                  */
 271                 sg_cpu->iowait_boost >>= 1;
 272                 if (sg_cpu->iowait_boost < IOWAIT_BOOST_MIN) {
 273                         sg_cpu->iowait_boost = 0;
 274                         return;
 275                 }
 276         }
 277
 278         sg_cpu->iowait_boost_pending = false;
 279
 280         /*
 281          * sg_cpu->util is already in capacity scale; convert iowait_boost
 282          * into the same scale so we can compare.
 283          */
 284         boost = (sg_cpu->iowait_boost * max_cap) >> SCHED_CAPACITY_SHIFT;
 285         boost = uclamp_rq_util_with(cpu_rq(sg_cpu->cpu), boost, NULL);
 286         if (sg_cpu->util < boost)
 287                 sg_cpu->util = boost;
 288 }
 289
 290 #ifdef CONFIG_NO_HZ_COMMON
 291 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
 292 {
 293         unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
 294         bool ret = idle_calls == sg_cpu->saved_idle_calls;
 295
 296         sg_cpu->saved_idle_calls = idle_calls;
 297         return ret;
 298 }
 299 #else
 300 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
 301 #endif /* CONFIG_NO_HZ_COMMON */
 302
 303 /*
 304  * Make sugov_should_update_freq() ignore the rate limit when DL
 305  * has increased the utilization.
 306  */
 307 static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu)
 308 {
 309         if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_dl)
 310                 sg_cpu->sg_policy->limits_changed = true;
 311 }
 312
 313 static inline bool sugov_update_single_common(struct sugov_cpu *sg_cpu,
 314                                               u64 time, unsigned long max_cap,
 315                                               unsigned int flags)
 316 {
 317         sugov_iowait_boost(sg_cpu, time, flags);
 318         sg_cpu->last_update = time;
 319
 320         ignore_dl_rate_limit(sg_cpu);
 321
 322         if (!sugov_should_update_freq(sg_cpu->sg_policy, time))
 323                 return false;
 324
 325         sugov_get_util(sg_cpu);
 326         sugov_iowait_apply(sg_cpu, time, max_cap);
 327
 328         return true;
 329 }
 330
 331 static void sugov_update_single_freq(struct update_util_data *hook, u64 time,
 332                                      unsigned int flags)
 333 {
 334         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 335         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 336         unsigned int cached_freq = sg_policy->cached_raw_freq;
 337         unsigned long max_cap;
 338         unsigned int next_f;
 339
 340         max_cap = arch_scale_cpu_capacity(sg_cpu->cpu);
 341
 342         if (!sugov_update_single_common(sg_cpu, time, max_cap, flags))
 343                 return;
 344
 345         next_f = get_next_freq(sg_policy, sg_cpu->util, max_cap);
 346         /*
 347          * Do not reduce the frequency if the CPU has not been idle
 348          * recently, as the reduction is likely to be premature then.
 349          *
 350          * Except when the rq is capped by uclamp_max.
 351          */
 352         if (!uclamp_rq_is_capped(cpu_rq(sg_cpu->cpu)) &&
 353             sugov_cpu_is_busy(sg_cpu) && next_f < sg_policy->next_freq &&
 354             !sg_policy->need_freq_update) {
 355                 next_f = sg_policy->next_freq;
 356
 357                 /* Restore cached freq as next_freq has changed */
 358                 sg_policy->cached_raw_freq = cached_freq;
 359         }
 360
 361         if (!sugov_update_next_freq(sg_policy, time, next_f))
 362                 return;
 363
 364         /*
 365          * This code runs under rq->lock for the target CPU, so it won't run
 366          * concurrently on two different CPUs for the same target and it is not
 367          * necessary to acquire the lock in the fast switch case.
 368          */
 369         if (sg_policy->policy->fast_switch_enabled) {
 370                 cpufreq_driver_fast_switch(sg_policy->policy, next_f);
 371         } else {
 372                 raw_spin_lock(&sg_policy->update_lock);
 373                 sugov_deferred_update(sg_policy);
 374                 raw_spin_unlock(&sg_policy->update_lock);
 375         }
 376 }
 377
 378 static void sugov_update_single_perf(struct update_util_data *hook, u64 time,
 379                                      unsigned int flags)
 380 {
 381         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 382         unsigned long prev_util = sg_cpu->util;
 383         unsigned long max_cap;
 384
 385         /*
 386          * Fall back to the "frequency" path if frequency invariance is not
 387          * supported, because the direct mapping between the utilization and
 388          * the performance levels depends on the frequency invariance.
 389          */
 390         if (!arch_scale_freq_invariant()) {
 391                 sugov_update_single_freq(hook, time, flags);
 392                 return;
 393         }
 394
 395         max_cap = arch_scale_cpu_capacity(sg_cpu->cpu);
 396
 397         if (!sugov_update_single_common(sg_cpu, time, max_cap, flags))
 398                 return;
 399
 400         /*
 401          * Do not reduce the target performance level if the CPU has not been
 402          * idle recently, as the reduction is likely to be premature then.
 403          *
 404          * Except when the rq is capped by uclamp_max.
 405          */
 406         if (!uclamp_rq_is_capped(cpu_rq(sg_cpu->cpu)) &&
 407             sugov_cpu_is_busy(sg_cpu) && sg_cpu->util < prev_util)
 408                 sg_cpu->util = prev_util;
 409
 410         cpufreq_driver_adjust_perf(sg_cpu->cpu, map_util_perf(sg_cpu->bw_dl),
 411                                    map_util_perf(sg_cpu->util), max_cap);
 412
 413         sg_cpu->sg_policy->last_freq_update_time = time;
 414 }
 415
 416 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
 417 {
 418         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 419         struct cpufreq_policy *policy = sg_policy->policy;
 420         unsigned long util = 0, max_cap;
 421         unsigned int j;
 422
 423         max_cap = arch_scale_cpu_capacity(sg_cpu->cpu);
 424
 425         for_each_cpu(j, policy->cpus) {
 426                 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
 427
 428                 sugov_get_util(j_sg_cpu);
 429                 sugov_iowait_apply(j_sg_cpu, time, max_cap);
 430
 431                 util = max(j_sg_cpu->util, util);
 432         }
 433
 434         return get_next_freq(sg_policy, util, max_cap);
 435 }
 436
 437 static void
 438 sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
 439 {
 440         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 441         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 442         unsigned int next_f;
 443
 444         raw_spin_lock(&sg_policy->update_lock);
 445
 446         sugov_iowait_boost(sg_cpu, time, flags);
 447         sg_cpu->last_update = time;
 448
 449         ignore_dl_rate_limit(sg_cpu);
 450
 451         if (sugov_should_update_freq(sg_policy, time)) {
 452                 next_f = sugov_next_freq_shared(sg_cpu, time);
 453
 454                 if (!sugov_update_next_freq(sg_policy, time, next_f))
 455                         goto unlock;
 456
 457                 if (sg_policy->policy->fast_switch_enabled)
 458                         cpufreq_driver_fast_switch(sg_policy->policy, next_f);
 459                 else
 460                         sugov_deferred_update(sg_policy);
 461         }
 462 unlock:
 463         raw_spin_unlock(&sg_policy->update_lock);
 464 }
 465
 466 static void sugov_work(struct kthread_work *work)
 467 {
 468         struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
 469         unsigned int freq;
 470         unsigned long flags;
 471
 472         /*
 473          * Hold sg_policy->update_lock shortly to handle the case where:
 474          * in case sg_policy->next_freq is read here, and then updated by
 475          * sugov_deferred_update() just before work_in_progress is set to false
 476          * here, we may miss queueing the new update.
 477          *
 478          * Note: If a work was queued after the update_lock is released,
 479          * sugov_work() will just be called again by kthread_work code; and the
 480          * request will be proceed before the sugov thread sleeps.
 481          */
 482         raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
 483         freq = sg_policy->next_freq;
 484         sg_policy->work_in_progress = false;
 485         raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
 486
 487         mutex_lock(&sg_policy->work_lock);
 488         __cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
 489         mutex_unlock(&sg_policy->work_lock);
 490 }
 491
 492 static void sugov_irq_work(struct irq_work *irq_work)
 493 {
 494         struct sugov_policy *sg_policy;
 495
 496         sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
 497
 498         kthread_queue_work(&sg_policy->worker, &sg_policy->work);
 499 }
 500
 501 /************************** sysfs interface ************************/
 502
 503 static struct sugov_tunables *global_tunables;
 504 static DEFINE_MUTEX(global_tunables_lock);
 505
 506 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
 507 {
 508         return container_of(attr_set, struct sugov_tunables, attr_set);
 509 }
 510
 511 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
 512 {
 513         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 514
 515         return sprintf(buf, "%u\n", tunables->rate_limit_us);
 516 }
 517
 518 static ssize_t
 519 rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
 520 {
 521         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 522         struct sugov_policy *sg_policy;
 523         unsigned int rate_limit_us;
 524
 525         if (kstrtouint(buf, 10, &rate_limit_us))
 526                 return -EINVAL;
 527
 528         tunables->rate_limit_us = rate_limit_us;
 529
 530         list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
 531                 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
 532
 533         return count;
 534 }
 535
 536 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
 537
 538 static struct attribute *sugov_attrs[] = {
 539         &rate_limit_us.attr,
 540         NULL
 541 };
 542 ATTRIBUTE_GROUPS(sugov);
 543
 544 static void sugov_tunables_free(struct kobject *kobj)
 545 {
 546         struct gov_attr_set *attr_set = to_gov_attr_set(kobj);
 547
 548         kfree(to_sugov_tunables(attr_set));
 549 }
 550
 551 static const struct kobj_type sugov_tunables_ktype = {
 552         .default_groups = sugov_groups,
 553         .sysfs_ops = &governor_sysfs_ops,
 554         .release = &sugov_tunables_free,
 555 };
 556
 557 /********************** cpufreq governor interface *********************/
 558
 559 struct cpufreq_governor schedutil_gov;
 560
 561 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
 562 {
 563         struct sugov_policy *sg_policy;
 564
 565         sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
 566         if (!sg_policy)
 567                 return NULL;
 568
 569         sg_policy->policy = policy;
 570         raw_spin_lock_init(&sg_policy->update_lock);
 571         return sg_policy;
 572 }
 573
 574 static void sugov_policy_free(struct sugov_policy *sg_policy)
 575 {
 576         kfree(sg_policy);
 577 }
 578
 579 static int sugov_kthread_create(struct sugov_policy *sg_policy)
 580 {
 581         struct task_struct *thread;
 582         struct sched_attr attr = {
 583                 .size           = sizeof(struct sched_attr),
 584                 .sched_policy   = SCHED_DEADLINE,
 585                 .sched_flags    = SCHED_FLAG_SUGOV,
 586                 .sched_nice     = 0,
 587                 .sched_priority = 0,
 588                 /*
 589                  * Fake (unused) bandwidth; workaround to "fix"
 590                  * priority inheritance.
 591                  */
 592                 .sched_runtime  =  1000000,
 593                 .sched_deadline = 10000000,
 594                 .sched_period   = 10000000,
 595         };
 596         struct cpufreq_policy *policy = sg_policy->policy;
 597         int ret;
 598
 599         /* kthread only required for slow path */
 600         if (policy->fast_switch_enabled)
 601                 return 0;
 602
 603         kthread_init_work(&sg_policy->work, sugov_work);
 604         kthread_init_worker(&sg_policy->worker);
 605         thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
 606                                 "sugov:%d",
 607                                 cpumask_first(policy->related_cpus));
 608         if (IS_ERR(thread)) {
 609                 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
 610                 return PTR_ERR(thread);
 611         }
 612
 613         ret = sched_setattr_nocheck(thread, &attr);
 614         if (ret) {
 615                 kthread_stop(thread);
 616                 pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
 617                 return ret;
 618         }
 619
 620         sg_policy->thread = thread;
 621         kthread_bind_mask(thread, policy->related_cpus);
 622         init_irq_work(&sg_policy->irq_work, sugov_irq_work);
 623         mutex_init(&sg_policy->work_lock);
 624
 625         wake_up_process(thread);
 626
 627         return 0;
 628 }
 629
 630 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
 631 {
 632         /* kthread only required for slow path */
 633         if (sg_policy->policy->fast_switch_enabled)
 634                 return;
 635
 636         kthread_flush_worker(&sg_policy->worker);
 637         kthread_stop(sg_policy->thread);
 638         mutex_destroy(&sg_policy->work_lock);
 639 }
 640
 641 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
 642 {
 643         struct sugov_tunables *tunables;
 644
 645         tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
 646         if (tunables) {
 647                 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
 648                 if (!have_governor_per_policy())
 649                         global_tunables = tunables;
 650         }
 651         return tunables;
 652 }
 653
 654 static void sugov_clear_global_tunables(void)
 655 {
 656         if (!have_governor_per_policy())
 657                 global_tunables = NULL;
 658 }
 659
 660 static int sugov_init(struct cpufreq_policy *policy)
 661 {
 662         struct sugov_policy *sg_policy;
 663         struct sugov_tunables *tunables;
 664         int ret = 0;
 665
 666         /* State should be equivalent to EXIT */
 667         if (policy->governor_data)
 668                 return -EBUSY;
 669
 670         cpufreq_enable_fast_switch(policy);
 671
 672         sg_policy = sugov_policy_alloc(policy);
 673         if (!sg_policy) {
 674                 ret = -ENOMEM;
 675                 goto disable_fast_switch;
 676         }
 677
 678         ret = sugov_kthread_create(sg_policy);
 679         if (ret)
 680                 goto free_sg_policy;
 681
 682         mutex_lock(&global_tunables_lock);
 683
 684         if (global_tunables) {
 685                 if (WARN_ON(have_governor_per_policy())) {
 686                         ret = -EINVAL;
 687                         goto stop_kthread;
 688                 }
 689                 policy->governor_data = sg_policy;
 690                 sg_policy->tunables = global_tunables;
 691
 692                 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
 693                 goto out;
 694         }
 695
 696         tunables = sugov_tunables_alloc(sg_policy);
 697         if (!tunables) {
 698                 ret = -ENOMEM;
 699                 goto stop_kthread;
 700         }
 701
 702         tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
 703
 704         policy->governor_data = sg_policy;
 705         sg_policy->tunables = tunables;
 706
 707         ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
 708                                    get_governor_parent_kobj(policy), "%s",
 709                                    schedutil_gov.name);
 710         if (ret)
 711                 goto fail;
 712
 713 out:
 714         mutex_unlock(&global_tunables_lock);
 715         return 0;
 716
 717 fail:
 718         kobject_put(&tunables->attr_set.kobj);
 719         policy->governor_data = NULL;
 720         sugov_clear_global_tunables();
 721
 722 stop_kthread:
 723         sugov_kthread_stop(sg_policy);
 724         mutex_unlock(&global_tunables_lock);
 725
 726 free_sg_policy:
 727         sugov_policy_free(sg_policy);
 728
 729 disable_fast_switch:
 730         cpufreq_disable_fast_switch(policy);
 731
 732         pr_err("initialization failed (error %d)\n", ret);
 733         return ret;
 734 }
 735
 736 static void sugov_exit(struct cpufreq_policy *policy)
 737 {
 738         struct sugov_policy *sg_policy = policy->governor_data;
 739         struct sugov_tunables *tunables = sg_policy->tunables;
 740         unsigned int count;
 741
 742         mutex_lock(&global_tunables_lock);
 743
 744         count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
 745         policy->governor_data = NULL;
 746         if (!count)
 747                 sugov_clear_global_tunables();
 748
 749         mutex_unlock(&global_tunables_lock);
 750
 751         sugov_kthread_stop(sg_policy);
 752         sugov_policy_free(sg_policy);
 753         cpufreq_disable_fast_switch(policy);
 754 }
 755
 756 static int sugov_start(struct cpufreq_policy *policy)
 757 {
 758         struct sugov_policy *sg_policy = policy->governor_data;
 759         void (*uu)(struct update_util_data *data, u64 time, unsigned int flags);
 760         unsigned int cpu;
 761
 762         sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
 763         sg_policy->last_freq_update_time        = 0;
 764         sg_policy->next_freq                    = 0;
 765         sg_policy->work_in_progress             = false;
 766         sg_policy->limits_changed               = false;
 767         sg_policy->cached_raw_freq              = 0;
 768
 769         sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
 770
 771         for_each_cpu(cpu, policy->cpus) {
 772                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 773
 774                 memset(sg_cpu, 0, sizeof(*sg_cpu));
 775                 sg_cpu->cpu                     = cpu;
 776                 sg_cpu->sg_policy               = sg_policy;
 777         }
 778
 779         if (policy_is_shared(policy))
 780                 uu = sugov_update_shared;
 781         else if (policy->fast_switch_enabled && cpufreq_driver_has_adjust_perf())
 782                 uu = sugov_update_single_perf;
 783         else
 784                 uu = sugov_update_single_freq;
 785
 786         for_each_cpu(cpu, policy->cpus) {
 787                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 788
 789                 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util, uu);
 790         }
 791         return 0;
 792 }
 793
 794 static void sugov_stop(struct cpufreq_policy *policy)
 795 {
 796         struct sugov_policy *sg_policy = policy->governor_data;
 797         unsigned int cpu;
 798
 799         for_each_cpu(cpu, policy->cpus)
 800                 cpufreq_remove_update_util_hook(cpu);
 801
 802         synchronize_rcu();
 803
 804         if (!policy->fast_switch_enabled) {
 805                 irq_work_sync(&sg_policy->irq_work);
 806                 kthread_cancel_work_sync(&sg_policy->work);
 807         }
 808 }
 809
 810 static void sugov_limits(struct cpufreq_policy *policy)
 811 {
 812         struct sugov_policy *sg_policy = policy->governor_data;
 813
 814         if (!policy->fast_switch_enabled) {
 815                 mutex_lock(&sg_policy->work_lock);
 816                 cpufreq_policy_apply_limits(policy);
 817                 mutex_unlock(&sg_policy->work_lock);
 818         }
 819
 820         sg_policy->limits_changed = true;
 821 }
 822
 823 struct cpufreq_governor schedutil_gov = {
 824         .name                   = "schedutil",
 825         .owner                  = THIS_MODULE,
 826         .flags                  = CPUFREQ_GOV_DYNAMIC_SWITCHING,
 827         .init                   = sugov_init,
 828         .exit                   = sugov_exit,
 829         .start                  = sugov_start,
 830         .stop                   = sugov_stop,
 831         .limits                 = sugov_limits,
 832 };
 833
 834 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
 835 struct cpufreq_governor *cpufreq_default_governor(void)
 836 {
 837         return &schedutil_gov;
 838 }
 839 #endif
 840
 841 cpufreq_governor_init(schedutil_gov);
 842
 843 #ifdef CONFIG_ENERGY_MODEL
 844 static void rebuild_sd_workfn(struct work_struct *work)
 845 {
 846         rebuild_sched_domains_energy();
 847 }
 848 static DECLARE_WORK(rebuild_sd_work, rebuild_sd_workfn);
 849
 850 /*
 851  * EAS shouldn't be attempted without sugov, so rebuild the sched_domains
 852  * on governor changes to make sure the scheduler knows about it.
 853  */
 854 void sched_cpufreq_governor_change(struct cpufreq_policy *policy,
 855                                   struct cpufreq_governor *old_gov)
 856 {
 857         if (old_gov == &schedutil_gov || policy->governor == &schedutil_gov) {
 858                 /*
 859                  * When called from the cpufreq_register_driver() path, the
 860                  * cpu_hotplug_lock is already held, so use a work item to
 861                  * avoid nested locking in rebuild_sched_domains().
 862                  */
 863                 schedule_work(&rebuild_sd_work);
 864         }
 865
 866 }
 867 #endif