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