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 pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  11 #include "sched.h"
  12
  13 #include <linux/sched/cpufreq.h>
  14 #include <trace/events/power.h>
  15
  16 #define IOWAIT_BOOST_MIN        (SCHED_CAPACITY_SCALE / 8)
  17
  18 struct sugov_tunables {
  19         struct gov_attr_set     attr_set;
  20         unsigned int            rate_limit_us;
  21 };
  22
  23 struct sugov_policy {
  24         struct cpufreq_policy   *policy;
  25
  26         struct sugov_tunables   *tunables;
  27         struct list_head        tunables_hook;
  28
  29         raw_spinlock_t          update_lock;
  30         u64                     last_freq_update_time;
  31         s64                     freq_update_delay_ns;
  32         unsigned int            next_freq;
  33         unsigned int            cached_raw_freq;
  34
  35         /* The next fields are only needed if fast switch cannot be used: */
  36         struct                  irq_work irq_work;
  37         struct                  kthread_work work;
  38         struct                  mutex work_lock;
  39         struct                  kthread_worker worker;
  40         struct task_struct      *thread;
  41         bool                    work_in_progress;
  42
  43         bool                    limits_changed;
  44         bool                    need_freq_update;
  45 };
  46
  47 struct sugov_cpu {
  48         struct update_util_data update_util;
  49         struct sugov_policy     *sg_policy;
  50         unsigned int            cpu;
  51
  52         bool                    iowait_boost_pending;
  53         unsigned int            iowait_boost;
  54         u64                     last_update;
  55
  56         unsigned long           util;
  57         unsigned long           bw_dl;
  58         unsigned long           max;
  59
  60         /* The field below is for single-CPU policies only: */
  61 #ifdef CONFIG_NO_HZ_COMMON
  62         unsigned long           saved_idle_calls;
  63 #endif
  64 };
  65
  66 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
  67
  68 /************************ Governor internals ***********************/
  69
  70 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
  71 {
  72         s64 delta_ns;
  73
  74         /*
  75          * Since cpufreq_update_util() is called with rq->lock held for
  76          * the @target_cpu, our per-CPU data is fully serialized.
  77          *
  78          * However, drivers cannot in general deal with cross-CPU
  79          * requests, so while get_next_freq() will work, our
  80          * sugov_update_commit() call may not for the fast switching platforms.
  81          *
  82          * Hence stop here for remote requests if they aren't supported
  83          * by the hardware, as calculating the frequency is pointless if
  84          * we cannot in fact act on it.
  85          *
  86          * This is needed on the slow switching platforms too to prevent CPUs
  87          * going offline from leaving stale IRQ work items behind.
  88          */
  89         if (!cpufreq_this_cpu_can_update(sg_policy->policy))
  90                 return false;
  91
  92         if (unlikely(sg_policy->limits_changed)) {
  93                 sg_policy->limits_changed = false;
  94                 sg_policy->need_freq_update = true;
  95                 return true;
  96         }
  97
  98         delta_ns = time - sg_policy->last_freq_update_time;
  99
 100         return delta_ns >= sg_policy->freq_update_delay_ns;
 101 }
 102
 103 static bool sugov_update_next_freq(struct sugov_policy *sg_policy, u64 time,
 104                                    unsigned int next_freq)
 105 {
 106         if (sg_policy->need_freq_update)
 107                 sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
 108         else if (sg_policy->next_freq == next_freq)
 109                 return false;
 110
 111         sg_policy->next_freq = next_freq;
 112         sg_policy->last_freq_update_time = time;
 113
 114         return true;
 115 }
 116
 117 static void sugov_deferred_update(struct sugov_policy *sg_policy)
 118 {
 119         if (!sg_policy->work_in_progress) {
 120                 sg_policy->work_in_progress = true;
 121                 irq_work_queue(&sg_policy->irq_work);
 122         }
 123 }
 124
 125 /**
 126  * get_next_freq - Compute a new frequency for a given cpufreq policy.
 127  * @sg_policy: schedutil policy object to compute the new frequency for.
 128  * @util: Current CPU utilization.
 129  * @max: CPU capacity.
 130  *
 131  * If the utilization is frequency-invariant, choose the new frequency to be
 132  * proportional to it, that is
 133  *
 134  * next_freq = C * max_freq * util / max
 135  *
 136  * Otherwise, approximate the would-be frequency-invariant utilization by
 137  * util_raw * (curr_freq / max_freq) which leads to
 138  *
 139  * next_freq = C * curr_freq * util_raw / max
 140  *
 141  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
 142  *
 143  * The lowest driver-supported frequency which is equal or greater than the raw
 144  * next_freq (as calculated above) is returned, subject to policy min/max and
 145  * cpufreq driver limitations.
 146  */
 147 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
 148                                   unsigned long util, unsigned long max)
 149 {
 150         struct cpufreq_policy *policy = sg_policy->policy;
 151         unsigned int freq = arch_scale_freq_invariant() ?
 152                                 policy->cpuinfo.max_freq : policy->cur;
 153
 154         util = map_util_perf(util);
 155         freq = map_util_freq(util, freq, max);
 156
 157         if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
 158                 return sg_policy->next_freq;
 159
 160         sg_policy->cached_raw_freq = freq;
 161         return cpufreq_driver_resolve_freq(policy, freq);
 162 }
 163
 164 static void sugov_get_util(struct sugov_cpu *sg_cpu)
 165 {
 166         struct rq *rq = cpu_rq(sg_cpu->cpu);
 167         unsigned long max = arch_scale_cpu_capacity(sg_cpu->cpu);
 168
 169         sg_cpu->max = max;
 170         sg_cpu->bw_dl = cpu_bw_dl(rq);
 171         sg_cpu->util = effective_cpu_util(sg_cpu->cpu, cpu_util_cfs(sg_cpu->cpu), max,
 172                                           FREQUENCY_UTIL, NULL);
 173 }
 174
 175 /**
 176  * sugov_iowait_reset() - Reset the IO boost status of a CPU.
 177  * @sg_cpu: the sugov data for the CPU to boost
 178  * @time: the update time from the caller
 179  * @set_iowait_boost: true if an IO boost has been requested
 180  *
 181  * The IO wait boost of a task is disabled after a tick since the last update
 182  * of a CPU. If a new IO wait boost is requested after more then a tick, then
 183  * we enable the boost starting from IOWAIT_BOOST_MIN, which improves energy
 184  * efficiency by ignoring sporadic wakeups from IO.
 185  */
 186 static bool sugov_iowait_reset(struct sugov_cpu *sg_cpu, u64 time,
 187                                bool set_iowait_boost)
 188 {
 189         s64 delta_ns = time - sg_cpu->last_update;
 190
 191         /* Reset boost only if a tick has elapsed since last request */
 192         if (delta_ns <= TICK_NSEC)
 193                 return false;
 194
 195         sg_cpu->iowait_boost = set_iowait_boost ? IOWAIT_BOOST_MIN : 0;
 196         sg_cpu->iowait_boost_pending = set_iowait_boost;
 197
 198         return true;
 199 }
 200
 201 /**
 202  * sugov_iowait_boost() - Updates the IO boost status of a CPU.
 203  * @sg_cpu: the sugov data for the CPU to boost
 204  * @time: the update time from the caller
 205  * @flags: SCHED_CPUFREQ_IOWAIT if the task is waking up after an IO wait
 206  *
 207  * Each time a task wakes up after an IO operation, the CPU utilization can be
 208  * boosted to a certain utilization which doubles at each "frequent and
 209  * successive" wakeup from IO, ranging from IOWAIT_BOOST_MIN to the utilization
 210  * of the maximum OPP.
 211  *
 212  * To keep doubling, an IO boost has to be requested at least once per tick,
 213  * otherwise we restart from the utilization of the minimum OPP.
 214  */
 215 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
 216                                unsigned int flags)
 217 {
 218         bool set_iowait_boost = flags & SCHED_CPUFREQ_IOWAIT;
 219
 220         /* Reset boost if the CPU appears to have been idle enough */
 221         if (sg_cpu->iowait_boost &&
 222             sugov_iowait_reset(sg_cpu, time, set_iowait_boost))
 223                 return;
 224
 225         /* Boost only tasks waking up after IO */
 226         if (!set_iowait_boost)
 227                 return;
 228
 229         /* Ensure boost doubles only one time at each request */
 230         if (sg_cpu->iowait_boost_pending)
 231                 return;
 232         sg_cpu->iowait_boost_pending = true;
 233
 234         /* Double the boost at each request */
 235         if (sg_cpu->iowait_boost) {
 236                 sg_cpu->iowait_boost =
 237                         min_t(unsigned int, sg_cpu->iowait_boost << 1, SCHED_CAPACITY_SCALE);
 238                 return;
 239         }
 240
 241         /* First wakeup after IO: start with minimum boost */
 242         sg_cpu->iowait_boost = IOWAIT_BOOST_MIN;
 243 }
 244
 245 /**
 246  * sugov_iowait_apply() - Apply the IO boost to a CPU.
 247  * @sg_cpu: the sugov data for the cpu to boost
 248  * @time: the update time from the caller
 249  *
 250  * A CPU running a task which woken up after an IO operation can have its
 251  * utilization boosted to speed up the completion of those IO operations.
 252  * The IO boost value is increased each time a task wakes up from IO, in
 253  * sugov_iowait_apply(), and it's instead decreased by this function,
 254  * each time an increase has not been requested (!iowait_boost_pending).
 255  *
 256  * A CPU which also appears to have been idle for at least one tick has also
 257  * its IO boost utilization reset.
 258  *
 259  * This mechanism is designed to boost high frequently IO waiting tasks, while
 260  * being more conservative on tasks which does sporadic IO operations.
 261  */
 262 static void sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time)
 263 {
 264         unsigned long boost;
 265
 266         /* No boost currently required */
 267         if (!sg_cpu->iowait_boost)
 268                 return;
 269
 270         /* Reset boost if the CPU appears to have been idle enough */
 271         if (sugov_iowait_reset(sg_cpu, time, false))
 272                 return;
 273
 274         if (!sg_cpu->iowait_boost_pending) {
 275                 /*
 276                  * No boost pending; reduce the boost value.
 277                  */
 278                 sg_cpu->iowait_boost >>= 1;
 279                 if (sg_cpu->iowait_boost < IOWAIT_BOOST_MIN) {
 280                         sg_cpu->iowait_boost = 0;
 281                         return;
 282                 }
 283         }
 284
 285         sg_cpu->iowait_boost_pending = false;
 286
 287         /*
 288          * sg_cpu->util is already in capacity scale; convert iowait_boost
 289          * into the same scale so we can compare.
 290          */
 291         boost = (sg_cpu->iowait_boost * sg_cpu->max) >> SCHED_CAPACITY_SHIFT;
 292         boost = uclamp_rq_util_with(cpu_rq(sg_cpu->cpu), boost, NULL);
 293         if (sg_cpu->util < boost)
 294                 sg_cpu->util = boost;
 295 }
 296
 297 #ifdef CONFIG_NO_HZ_COMMON
 298 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
 299 {
 300         unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
 301         bool ret = idle_calls == sg_cpu->saved_idle_calls;
 302
 303         sg_cpu->saved_idle_calls = idle_calls;
 304         return ret;
 305 }
 306 #else
 307 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
 308 #endif /* CONFIG_NO_HZ_COMMON */
 309
 310 /*
 311  * Make sugov_should_update_freq() ignore the rate limit when DL
 312  * has increased the utilization.
 313  */
 314 static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu)
 315 {
 316         if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_dl)
 317                 sg_cpu->sg_policy->limits_changed = true;
 318 }
 319
 320 static inline bool sugov_update_single_common(struct sugov_cpu *sg_cpu,
 321                                               u64 time, unsigned int flags)
 322 {
 323         sugov_iowait_boost(sg_cpu, time, flags);
 324         sg_cpu->last_update = time;
 325
 326         ignore_dl_rate_limit(sg_cpu);
 327
 328         if (!sugov_should_update_freq(sg_cpu->sg_policy, time))
 329                 return false;
 330
 331         sugov_get_util(sg_cpu);
 332         sugov_iowait_apply(sg_cpu, time);
 333
 334         return true;
 335 }
 336
 337 static void sugov_update_single_freq(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 int cached_freq = sg_policy->cached_raw_freq;
 343         unsigned int next_f;
 344
 345         if (!sugov_update_single_common(sg_cpu, time, flags))
 346                 return;
 347
 348         next_f = get_next_freq(sg_policy, sg_cpu->util, sg_cpu->max);
 349         /*
 350          * Do not reduce the frequency if the CPU has not been idle
 351          * recently, as the reduction is likely to be premature then.
 352          *
 353          * Except when the rq is capped by uclamp_max.
 354          */
 355         if (!uclamp_rq_is_capped(cpu_rq(sg_cpu->cpu)) &&
 356             sugov_cpu_is_busy(sg_cpu) && next_f < sg_policy->next_freq) {
 357                 next_f = sg_policy->next_freq;
 358
 359                 /* Restore cached freq as next_freq has changed */
 360                 sg_policy->cached_raw_freq = cached_freq;
 361         }
 362
 363         if (!sugov_update_next_freq(sg_policy, time, next_f))
 364                 return;
 365
 366         /*
 367          * This code runs under rq->lock for the target CPU, so it won't run
 368          * concurrently on two different CPUs for the same target and it is not
 369          * necessary to acquire the lock in the fast switch case.
 370          */
 371         if (sg_policy->policy->fast_switch_enabled) {
 372                 cpufreq_driver_fast_switch(sg_policy->policy, next_f);
 373         } else {
 374                 raw_spin_lock(&sg_policy->update_lock);
 375                 sugov_deferred_update(sg_policy);
 376                 raw_spin_unlock(&sg_policy->update_lock);
 377         }
 378 }
 379
 380 static void sugov_update_single_perf(struct update_util_data *hook, u64 time,
 381                                      unsigned int flags)
 382 {
 383         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 384         unsigned long prev_util = sg_cpu->util;
 385
 386         /*
 387          * Fall back to the "frequency" path if frequency invariance is not
 388          * supported, because the direct mapping between the utilization and
 389          * the performance levels depends on the frequency invariance.
 390          */
 391         if (!arch_scale_freq_invariant()) {
 392                 sugov_update_single_freq(hook, time, flags);
 393                 return;
 394         }
 395
 396         if (!sugov_update_single_common(sg_cpu, time, flags))
 397                 return;
 398
 399         /*
 400          * Do not reduce the target performance level if the CPU has not been
 401          * idle recently, as the reduction is likely to be premature then.
 402          *
 403          * Except when the rq is capped by uclamp_max.
 404          */
 405         if (!uclamp_rq_is_capped(cpu_rq(sg_cpu->cpu)) &&
 406             sugov_cpu_is_busy(sg_cpu) && sg_cpu->util < prev_util)
 407                 sg_cpu->util = prev_util;
 408
 409         cpufreq_driver_adjust_perf(sg_cpu->cpu, map_util_perf(sg_cpu->bw_dl),
 410                                    map_util_perf(sg_cpu->util), sg_cpu->max);
 411
 412         sg_cpu->sg_policy->last_freq_update_time = time;
 413 }
 414
 415 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
 416 {
 417         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 418         struct cpufreq_policy *policy = sg_policy->policy;
 419         unsigned long util = 0, max = 1;
 420         unsigned int j;
 421
 422         for_each_cpu(j, policy->cpus) {
 423                 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
 424                 unsigned long j_util, j_max;
 425
 426                 sugov_get_util(j_sg_cpu);
 427                 sugov_iowait_apply(j_sg_cpu, time);
 428                 j_util = j_sg_cpu->util;
 429                 j_max = j_sg_cpu->max;
 430
 431                 if (j_util * max > j_max * util) {
 432                         util = j_util;
 433                         max = j_max;
 434                 }
 435         }
 436
 437         return get_next_freq(sg_policy, util, max);
 438 }
 439
 440 static void
 441 sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
 442 {
 443         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 444         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 445         unsigned int next_f;
 446
 447         raw_spin_lock(&sg_policy->update_lock);
 448
 449         sugov_iowait_boost(sg_cpu, time, flags);
 450         sg_cpu->last_update = time;
 451
 452         ignore_dl_rate_limit(sg_cpu);
 453
 454         if (sugov_should_update_freq(sg_policy, time)) {
 455                 next_f = sugov_next_freq_shared(sg_cpu, time);
 456
 457                 if (!sugov_update_next_freq(sg_policy, time, next_f))
 458                         goto unlock;
 459
 460                 if (sg_policy->policy->fast_switch_enabled)
 461                         cpufreq_driver_fast_switch(sg_policy->policy, next_f);
 462                 else
 463                         sugov_deferred_update(sg_policy);
 464         }
 465 unlock:
 466         raw_spin_unlock(&sg_policy->update_lock);
 467 }
 468
 469 static void sugov_work(struct kthread_work *work)
 470 {
 471         struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
 472         unsigned int freq;
 473         unsigned long flags;
 474
 475         /*
 476          * Hold sg_policy->update_lock shortly to handle the case where:
 477          * in case sg_policy->next_freq is read here, and then updated by
 478          * sugov_deferred_update() just before work_in_progress is set to false
 479          * here, we may miss queueing the new update.
 480          *
 481          * Note: If a work was queued after the update_lock is released,
 482          * sugov_work() will just be called again by kthread_work code; and the
 483          * request will be proceed before the sugov thread sleeps.
 484          */
 485         raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
 486         freq = sg_policy->next_freq;
 487         sg_policy->work_in_progress = false;
 488         raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
 489
 490         mutex_lock(&sg_policy->work_lock);
 491         __cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
 492         mutex_unlock(&sg_policy->work_lock);
 493 }
 494
 495 static void sugov_irq_work(struct irq_work *irq_work)
 496 {
 497         struct sugov_policy *sg_policy;
 498
 499         sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
 500
 501         kthread_queue_work(&sg_policy->worker, &sg_policy->work);
 502 }
 503
 504 /************************** sysfs interface ************************/
 505
 506 static struct sugov_tunables *global_tunables;
 507 static DEFINE_MUTEX(global_tunables_lock);
 508
 509 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
 510 {
 511         return container_of(attr_set, struct sugov_tunables, attr_set);
 512 }
 513
 514 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
 515 {
 516         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 517
 518         return sprintf(buf, "%u\n", tunables->rate_limit_us);
 519 }
 520
 521 static ssize_t
 522 rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
 523 {
 524         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 525         struct sugov_policy *sg_policy;
 526         unsigned int rate_limit_us;
 527
 528         if (kstrtouint(buf, 10, &rate_limit_us))
 529                 return -EINVAL;
 530
 531         tunables->rate_limit_us = rate_limit_us;
 532
 533         list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
 534                 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
 535
 536         return count;
 537 }
 538
 539 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
 540
 541 static struct attribute *sugov_attrs[] = {
 542         &rate_limit_us.attr,
 543         NULL
 544 };
 545 ATTRIBUTE_GROUPS(sugov);
 546
 547 static void sugov_tunables_free(struct kobject *kobj)
 548 {
 549         struct gov_attr_set *attr_set = container_of(kobj, struct gov_attr_set, kobj);
 550
 551         kfree(to_sugov_tunables(attr_set));
 552 }
 553
 554 static struct kobj_type sugov_tunables_ktype = {
 555         .default_groups = sugov_groups,
 556         .sysfs_ops = &governor_sysfs_ops,
 557         .release = &sugov_tunables_free,
 558 };
 559
 560 /********************** cpufreq governor interface *********************/
 561
 562 struct cpufreq_governor schedutil_gov;
 563
 564 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
 565 {
 566         struct sugov_policy *sg_policy;
 567
 568         sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
 569         if (!sg_policy)
 570                 return NULL;
 571
 572         sg_policy->policy = policy;
 573         raw_spin_lock_init(&sg_policy->update_lock);
 574         return sg_policy;
 575 }
 576
 577 static void sugov_policy_free(struct sugov_policy *sg_policy)
 578 {
 579         kfree(sg_policy);
 580 }
 581
 582 static int sugov_kthread_create(struct sugov_policy *sg_policy)
 583 {
 584         struct task_struct *thread;
 585         struct sched_attr attr = {
 586                 .size           = sizeof(struct sched_attr),
 587                 .sched_policy   = SCHED_DEADLINE,
 588                 .sched_flags    = SCHED_FLAG_SUGOV,
 589                 .sched_nice     = 0,
 590                 .sched_priority = 0,
 591                 /*
 592                  * Fake (unused) bandwidth; workaround to "fix"
 593                  * priority inheritance.
 594                  */
 595                 .sched_runtime  =  1000000,
 596                 .sched_deadline = 10000000,
 597                 .sched_period   = 10000000,
 598         };
 599         struct cpufreq_policy *policy = sg_policy->policy;
 600         int ret;
 601
 602         /* kthread only required for slow path */
 603         if (policy->fast_switch_enabled)
 604                 return 0;
 605
 606         kthread_init_work(&sg_policy->work, sugov_work);
 607         kthread_init_worker(&sg_policy->worker);
 608         thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
 609                                 "sugov:%d",
 610                                 cpumask_first(policy->related_cpus));
 611         if (IS_ERR(thread)) {
 612                 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
 613                 return PTR_ERR(thread);
 614         }
 615
 616         ret = sched_setattr_nocheck(thread, &attr);
 617         if (ret) {
 618                 kthread_stop(thread);
 619                 pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
 620                 return ret;
 621         }
 622
 623         sg_policy->thread = thread;
 624         kthread_bind_mask(thread, policy->related_cpus);
 625         init_irq_work(&sg_policy->irq_work, sugov_irq_work);
 626         mutex_init(&sg_policy->work_lock);
 627
 628         wake_up_process(thread);
 629
 630         return 0;
 631 }
 632
 633 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
 634 {
 635         /* kthread only required for slow path */
 636         if (sg_policy->policy->fast_switch_enabled)
 637                 return;
 638
 639         kthread_flush_worker(&sg_policy->worker);
 640         kthread_stop(sg_policy->thread);
 641         mutex_destroy(&sg_policy->work_lock);
 642 }
 643
 644 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
 645 {
 646         struct sugov_tunables *tunables;
 647
 648         tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
 649         if (tunables) {
 650                 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
 651                 if (!have_governor_per_policy())
 652                         global_tunables = tunables;
 653         }
 654         return tunables;
 655 }
 656
 657 static void sugov_clear_global_tunables(void)
 658 {
 659         if (!have_governor_per_policy())
 660                 global_tunables = NULL;
 661 }
 662
 663 static int sugov_init(struct cpufreq_policy *policy)
 664 {
 665         struct sugov_policy *sg_policy;
 666         struct sugov_tunables *tunables;
 667         int ret = 0;
 668
 669         /* State should be equivalent to EXIT */
 670         if (policy->governor_data)
 671                 return -EBUSY;
 672
 673         cpufreq_enable_fast_switch(policy);
 674
 675         sg_policy = sugov_policy_alloc(policy);
 676         if (!sg_policy) {
 677                 ret = -ENOMEM;
 678                 goto disable_fast_switch;
 679         }
 680
 681         ret = sugov_kthread_create(sg_policy);
 682         if (ret)
 683                 goto free_sg_policy;
 684
 685         mutex_lock(&global_tunables_lock);
 686
 687         if (global_tunables) {
 688                 if (WARN_ON(have_governor_per_policy())) {
 689                         ret = -EINVAL;
 690                         goto stop_kthread;
 691                 }
 692                 policy->governor_data = sg_policy;
 693                 sg_policy->tunables = global_tunables;
 694
 695                 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
 696                 goto out;
 697         }
 698
 699         tunables = sugov_tunables_alloc(sg_policy);
 700         if (!tunables) {
 701                 ret = -ENOMEM;
 702                 goto stop_kthread;
 703         }
 704
 705         tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
 706
 707         policy->governor_data = sg_policy;
 708         sg_policy->tunables = tunables;
 709
 710         ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
 711                                    get_governor_parent_kobj(policy), "%s",
 712                                    schedutil_gov.name);
 713         if (ret)
 714                 goto fail;
 715
 716 out:
 717         mutex_unlock(&global_tunables_lock);
 718         return 0;
 719
 720 fail:
 721         kobject_put(&tunables->attr_set.kobj);
 722         policy->governor_data = NULL;
 723         sugov_clear_global_tunables();
 724
 725 stop_kthread:
 726         sugov_kthread_stop(sg_policy);
 727         mutex_unlock(&global_tunables_lock);
 728
 729 free_sg_policy:
 730         sugov_policy_free(sg_policy);
 731
 732 disable_fast_switch:
 733         cpufreq_disable_fast_switch(policy);
 734
 735         pr_err("initialization failed (error %d)\n", ret);
 736         return ret;
 737 }
 738
 739 static void sugov_exit(struct cpufreq_policy *policy)
 740 {
 741         struct sugov_policy *sg_policy = policy->governor_data;
 742         struct sugov_tunables *tunables = sg_policy->tunables;
 743         unsigned int count;
 744
 745         mutex_lock(&global_tunables_lock);
 746
 747         count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
 748         policy->governor_data = NULL;
 749         if (!count)
 750                 sugov_clear_global_tunables();
 751
 752         mutex_unlock(&global_tunables_lock);
 753
 754         sugov_kthread_stop(sg_policy);
 755         sugov_policy_free(sg_policy);
 756         cpufreq_disable_fast_switch(policy);
 757 }
 758
 759 static int sugov_start(struct cpufreq_policy *policy)
 760 {
 761         struct sugov_policy *sg_policy = policy->governor_data;
 762         void (*uu)(struct update_util_data *data, u64 time, unsigned int flags);
 763         unsigned int cpu;
 764
 765         sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
 766         sg_policy->last_freq_update_time        = 0;
 767         sg_policy->next_freq                    = 0;
 768         sg_policy->work_in_progress             = false;
 769         sg_policy->limits_changed               = false;
 770         sg_policy->cached_raw_freq              = 0;
 771
 772         sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
 773
 774         for_each_cpu(cpu, policy->cpus) {
 775                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 776
 777                 memset(sg_cpu, 0, sizeof(*sg_cpu));
 778                 sg_cpu->cpu                     = cpu;
 779                 sg_cpu->sg_policy               = sg_policy;
 780         }
 781
 782         if (policy_is_shared(policy))
 783                 uu = sugov_update_shared;
 784         else if (policy->fast_switch_enabled && cpufreq_driver_has_adjust_perf())
 785                 uu = sugov_update_single_perf;
 786         else
 787                 uu = sugov_update_single_freq;
 788
 789         for_each_cpu(cpu, policy->cpus) {
 790                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 791
 792                 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util, uu);
 793         }
 794         return 0;
 795 }
 796
 797 static void sugov_stop(struct cpufreq_policy *policy)
 798 {
 799         struct sugov_policy *sg_policy = policy->governor_data;
 800         unsigned int cpu;
 801
 802         for_each_cpu(cpu, policy->cpus)
 803                 cpufreq_remove_update_util_hook(cpu);
 804
 805         synchronize_rcu();
 806
 807         if (!policy->fast_switch_enabled) {
 808                 irq_work_sync(&sg_policy->irq_work);
 809                 kthread_cancel_work_sync(&sg_policy->work);
 810         }
 811 }
 812
 813 static void sugov_limits(struct cpufreq_policy *policy)
 814 {
 815         struct sugov_policy *sg_policy = policy->governor_data;
 816
 817         if (!policy->fast_switch_enabled) {
 818                 mutex_lock(&sg_policy->work_lock);
 819                 cpufreq_policy_apply_limits(policy);
 820                 mutex_unlock(&sg_policy->work_lock);
 821         }
 822
 823         sg_policy->limits_changed = true;
 824 }
 825
 826 struct cpufreq_governor schedutil_gov = {
 827         .name                   = "schedutil",
 828         .owner                  = THIS_MODULE,
 829         .flags                  = CPUFREQ_GOV_DYNAMIC_SWITCHING,
 830         .init                   = sugov_init,
 831         .exit                   = sugov_exit,
 832         .start                  = sugov_start,
 833         .stop                   = sugov_stop,
 834         .limits                 = sugov_limits,
 835 };
 836
 837 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
 838 struct cpufreq_governor *cpufreq_default_governor(void)
 839 {
 840         return &schedutil_gov;
 841 }
 842 #endif
 843
 844 cpufreq_governor_init(schedutil_gov);
 845
 846 #ifdef CONFIG_ENERGY_MODEL
 847 static void rebuild_sd_workfn(struct work_struct *work)
 848 {
 849         rebuild_sched_domains_energy();
 850 }
 851 static DECLARE_WORK(rebuild_sd_work, rebuild_sd_workfn);
 852
 853 /*
 854  * EAS shouldn't be attempted without sugov, so rebuild the sched_domains
 855  * on governor changes to make sure the scheduler knows about it.
 856  */
 857 void sched_cpufreq_governor_change(struct cpufreq_policy *policy,
 858                                   struct cpufreq_governor *old_gov)
 859 {
 860         if (old_gov == &schedutil_gov || policy->governor == &schedutil_gov) {
 861                 /*
 862                  * When called from the cpufreq_register_driver() path, the
 863                  * cpu_hotplug_lock is already held, so use a work item to
 864                  * avoid nested locking in rebuild_sched_domains().
 865                  */
 866                 schedule_work(&rebuild_sd_work);
 867         }
 868
 869 }
 870 #endif