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