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_min;
  51
  52         /* The field below is for single-CPU policies only: */
  53 #ifdef CONFIG_NO_HZ_COMMON
  54         unsigned long           saved_idle_calls;
  55 #endif
  56 };
  57
  58 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
  59
  60 /************************ Governor internals ***********************/
  61
  62 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
  63 {
  64         s64 delta_ns;
  65
  66         /*
  67          * Since cpufreq_update_util() is called with rq->lock held for
  68          * the @target_cpu, our per-CPU data is fully serialized.
  69          *
  70          * However, drivers cannot in general deal with cross-CPU
  71          * requests, so while get_next_freq() will work, our
  72          * sugov_update_commit() call may not for the fast switching platforms.
  73          *
  74          * Hence stop here for remote requests if they aren't supported
  75          * by the hardware, as calculating the frequency is pointless if
  76          * we cannot in fact act on it.
  77          *
  78          * This is needed on the slow switching platforms too to prevent CPUs
  79          * going offline from leaving stale IRQ work items behind.
  80          */
  81         if (!cpufreq_this_cpu_can_update(sg_policy->policy))
  82                 return false;
  83
  84         if (unlikely(sg_policy->limits_changed)) {
  85                 sg_policy->limits_changed = false;
  86                 sg_policy->need_freq_update = true;
  87                 return true;
  88         }
  89
  90         delta_ns = time - sg_policy->last_freq_update_time;
  91
  92         return delta_ns >= sg_policy->freq_update_delay_ns;
  93 }
  94
  95 static bool sugov_update_next_freq(struct sugov_policy *sg_policy, u64 time,
  96                                    unsigned int next_freq)
  97 {
  98         if (sg_policy->need_freq_update)
  99                 sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
 100         else if (sg_policy->next_freq == next_freq)
 101                 return false;
 102
 103         sg_policy->next_freq = next_freq;
 104         sg_policy->last_freq_update_time = time;
 105
 106         return true;
 107 }
 108
 109 static void sugov_deferred_update(struct sugov_policy *sg_policy)
 110 {
 111         if (!sg_policy->work_in_progress) {
 112                 sg_policy->work_in_progress = true;
 113                 irq_work_queue(&sg_policy->irq_work);
 114         }
 115 }
 116
 117 /**
 118  * get_capacity_ref_freq - get the reference frequency that has been used to
 119  * correlate frequency and compute capacity for a given cpufreq policy. We use
 120  * the CPU managing it for the arch_scale_freq_ref() call in the function.
 121  * @policy: the cpufreq policy of the CPU in question.
 122  *
 123  * Return: the reference CPU frequency to compute a capacity.
 124  */
 125 static __always_inline
 126 unsigned long get_capacity_ref_freq(struct cpufreq_policy *policy)
 127 {
 128         unsigned int freq = arch_scale_freq_ref(policy->cpu);
 129
 130         if (freq)
 131                 return freq;
 132
 133         if (arch_scale_freq_invariant())
 134                 return policy->cpuinfo.max_freq;
 135
 136         return policy->cur;
 137 }
 138
 139 /**
 140  * get_next_freq - Compute a new frequency for a given cpufreq policy.
 141  * @sg_policy: schedutil policy object to compute the new frequency for.
 142  * @util: Current CPU utilization.
 143  * @max: CPU capacity.
 144  *
 145  * If the utilization is frequency-invariant, choose the new frequency to be
 146  * proportional to it, that is
 147  *
 148  * next_freq = C * max_freq * util / max
 149  *
 150  * Otherwise, approximate the would-be frequency-invariant utilization by
 151  * util_raw * (curr_freq / max_freq) which leads to
 152  *
 153  * next_freq = C * curr_freq * util_raw / max
 154  *
 155  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
 156  *
 157  * The lowest driver-supported frequency which is equal or greater than the raw
 158  * next_freq (as calculated above) is returned, subject to policy min/max and
 159  * cpufreq driver limitations.
 160  */
 161 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
 162                                   unsigned long util, unsigned long max)
 163 {
 164         struct cpufreq_policy *policy = sg_policy->policy;
 165         unsigned int freq;
 166
 167         freq = get_capacity_ref_freq(policy);
 168         freq = map_util_freq(util, freq, max);
 169
 170         if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
 171                 return sg_policy->next_freq;
 172
 173         sg_policy->cached_raw_freq = freq;
 174         return cpufreq_driver_resolve_freq(policy, freq);
 175 }
 176
 177 unsigned long sugov_effective_cpu_perf(int cpu, unsigned long actual,
 178                                  unsigned long min,
 179                                  unsigned long max)
 180 {
 181         /* Add dvfs headroom to actual utilization */
 182         actual = map_util_perf(actual);
 183         /* Actually we don't need to target the max performance */
 184         if (actual < max)
 185                 max = actual;
 186
 187         /*
 188          * Ensure at least minimum performance while providing more compute
 189          * capacity when possible.
 190          */
 191         return max(min, max);
 192 }
 193
 194 static void sugov_get_util(struct sugov_cpu *sg_cpu, unsigned long boost)
 195 {
 196         unsigned long min, max, util = cpu_util_cfs_boost(sg_cpu->cpu);
 197
 198         util = effective_cpu_util(sg_cpu->cpu, util, &min, &max);
 199         util = max(util, boost);
 200         sg_cpu->bw_min = min;
 201         sg_cpu->util = sugov_effective_cpu_perf(sg_cpu->cpu, util, min, max);
 202 }
 203
 204 /**
 205  * sugov_iowait_reset() - Reset the IO boost status of a CPU.
 206  * @sg_cpu: the sugov data for the CPU to boost
 207  * @time: the update time from the caller
 208  * @set_iowait_boost: true if an IO boost has been requested
 209  *
 210  * The IO wait boost of a task is disabled after a tick since the last update
 211  * of a CPU. If a new IO wait boost is requested after more then a tick, then
 212  * we enable the boost starting from IOWAIT_BOOST_MIN, which improves energy
 213  * efficiency by ignoring sporadic wakeups from IO.
 214  */
 215 static bool sugov_iowait_reset(struct sugov_cpu *sg_cpu, u64 time,
 216                                bool set_iowait_boost)
 217 {
 218         s64 delta_ns = time - sg_cpu->last_update;
 219
 220         /* Reset boost only if a tick has elapsed since last request */
 221         if (delta_ns <= TICK_NSEC)
 222                 return false;
 223
 224         sg_cpu->iowait_boost = set_iowait_boost ? IOWAIT_BOOST_MIN : 0;
 225         sg_cpu->iowait_boost_pending = set_iowait_boost;
 226
 227         return true;
 228 }
 229
 230 /**
 231  * sugov_iowait_boost() - Updates the IO boost status of a CPU.
 232  * @sg_cpu: the sugov data for the CPU to boost
 233  * @time: the update time from the caller
 234  * @flags: SCHED_CPUFREQ_IOWAIT if the task is waking up after an IO wait
 235  *
 236  * Each time a task wakes up after an IO operation, the CPU utilization can be
 237  * boosted to a certain utilization which doubles at each "frequent and
 238  * successive" wakeup from IO, ranging from IOWAIT_BOOST_MIN to the utilization
 239  * of the maximum OPP.
 240  *
 241  * To keep doubling, an IO boost has to be requested at least once per tick,
 242  * otherwise we restart from the utilization of the minimum OPP.
 243  */
 244 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
 245                                unsigned int flags)
 246 {
 247         bool set_iowait_boost = flags & SCHED_CPUFREQ_IOWAIT;
 248
 249         /* Reset boost if the CPU appears to have been idle enough */
 250         if (sg_cpu->iowait_boost &&
 251             sugov_iowait_reset(sg_cpu, time, set_iowait_boost))
 252                 return;
 253
 254         /* Boost only tasks waking up after IO */
 255         if (!set_iowait_boost)
 256                 return;
 257
 258         /* Ensure boost doubles only one time at each request */
 259         if (sg_cpu->iowait_boost_pending)
 260                 return;
 261         sg_cpu->iowait_boost_pending = true;
 262
 263         /* Double the boost at each request */
 264         if (sg_cpu->iowait_boost) {
 265                 sg_cpu->iowait_boost =
 266                         min_t(unsigned int, sg_cpu->iowait_boost << 1, SCHED_CAPACITY_SCALE);
 267                 return;
 268         }
 269
 270         /* First wakeup after IO: start with minimum boost */
 271         sg_cpu->iowait_boost = IOWAIT_BOOST_MIN;
 272 }
 273
 274 /**
 275  * sugov_iowait_apply() - Apply the IO boost to a CPU.
 276  * @sg_cpu: the sugov data for the cpu to boost
 277  * @time: the update time from the caller
 278  * @max_cap: the max CPU capacity
 279  *
 280  * A CPU running a task which woken up after an IO operation can have its
 281  * utilization boosted to speed up the completion of those IO operations.
 282  * The IO boost value is increased each time a task wakes up from IO, in
 283  * sugov_iowait_apply(), and it's instead decreased by this function,
 284  * each time an increase has not been requested (!iowait_boost_pending).
 285  *
 286  * A CPU which also appears to have been idle for at least one tick has also
 287  * its IO boost utilization reset.
 288  *
 289  * This mechanism is designed to boost high frequently IO waiting tasks, while
 290  * being more conservative on tasks which does sporadic IO operations.
 291  */
 292 static unsigned long sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
 293                                unsigned long max_cap)
 294 {
 295         /* No boost currently required */
 296         if (!sg_cpu->iowait_boost)
 297                 return 0;
 298
 299         /* Reset boost if the CPU appears to have been idle enough */
 300         if (sugov_iowait_reset(sg_cpu, time, false))
 301                 return 0;
 302
 303         if (!sg_cpu->iowait_boost_pending) {
 304                 /*
 305                  * No boost pending; reduce the boost value.
 306                  */
 307                 sg_cpu->iowait_boost >>= 1;
 308                 if (sg_cpu->iowait_boost < IOWAIT_BOOST_MIN) {
 309                         sg_cpu->iowait_boost = 0;
 310                         return 0;
 311                 }
 312         }
 313
 314         sg_cpu->iowait_boost_pending = false;
 315
 316         /*
 317          * sg_cpu->util is already in capacity scale; convert iowait_boost
 318          * into the same scale so we can compare.
 319          */
 320         return (sg_cpu->iowait_boost * max_cap) >> SCHED_CAPACITY_SHIFT;
 321 }
 322
 323 #ifdef CONFIG_NO_HZ_COMMON
 324 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
 325 {
 326         unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
 327         bool ret = idle_calls == sg_cpu->saved_idle_calls;
 328
 329         sg_cpu->saved_idle_calls = idle_calls;
 330         return ret;
 331 }
 332 #else
 333 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
 334 #endif /* CONFIG_NO_HZ_COMMON */
 335
 336 /*
 337  * Make sugov_should_update_freq() ignore the rate limit when DL
 338  * has increased the utilization.
 339  */
 340 static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu)
 341 {
 342         if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_min)
 343                 sg_cpu->sg_policy->limits_changed = true;
 344 }
 345
 346 static inline bool sugov_update_single_common(struct sugov_cpu *sg_cpu,
 347                                               u64 time, unsigned long max_cap,
 348                                               unsigned int flags)
 349 {
 350         unsigned long boost;
 351
 352         sugov_iowait_boost(sg_cpu, time, flags);
 353         sg_cpu->last_update = time;
 354
 355         ignore_dl_rate_limit(sg_cpu);
 356
 357         if (!sugov_should_update_freq(sg_cpu->sg_policy, time))
 358                 return false;
 359
 360         boost = sugov_iowait_apply(sg_cpu, time, max_cap);
 361         sugov_get_util(sg_cpu, boost);
 362
 363         return true;
 364 }
 365
 366 static void sugov_update_single_freq(struct update_util_data *hook, u64 time,
 367                                      unsigned int flags)
 368 {
 369         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 370         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 371         unsigned int cached_freq = sg_policy->cached_raw_freq;
 372         unsigned long max_cap;
 373         unsigned int next_f;
 374
 375         max_cap = arch_scale_cpu_capacity(sg_cpu->cpu);
 376
 377         if (!sugov_update_single_common(sg_cpu, time, max_cap, flags))
 378                 return;
 379
 380         next_f = get_next_freq(sg_policy, sg_cpu->util, max_cap);
 381         /*
 382          * Do not reduce the frequency if the CPU has not been idle
 383          * recently, as the reduction is likely to be premature then.
 384          *
 385          * Except when the rq is capped by uclamp_max.
 386          */
 387         if (!uclamp_rq_is_capped(cpu_rq(sg_cpu->cpu)) &&
 388             sugov_cpu_is_busy(sg_cpu) && next_f < sg_policy->next_freq &&
 389             !sg_policy->need_freq_update) {
 390                 next_f = sg_policy->next_freq;
 391
 392                 /* Restore cached freq as next_freq has changed */
 393                 sg_policy->cached_raw_freq = cached_freq;
 394         }
 395
 396         if (!sugov_update_next_freq(sg_policy, time, next_f))
 397                 return;
 398
 399         /*
 400          * This code runs under rq->lock for the target CPU, so it won't run
 401          * concurrently on two different CPUs for the same target and it is not
 402          * necessary to acquire the lock in the fast switch case.
 403          */
 404         if (sg_policy->policy->fast_switch_enabled) {
 405                 cpufreq_driver_fast_switch(sg_policy->policy, next_f);
 406         } else {
 407                 raw_spin_lock(&sg_policy->update_lock);
 408                 sugov_deferred_update(sg_policy);
 409                 raw_spin_unlock(&sg_policy->update_lock);
 410         }
 411 }
 412
 413 static void sugov_update_single_perf(struct update_util_data *hook, u64 time,
 414                                      unsigned int flags)
 415 {
 416         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 417         unsigned long prev_util = sg_cpu->util;
 418         unsigned long max_cap;
 419
 420         /*
 421          * Fall back to the "frequency" path if frequency invariance is not
 422          * supported, because the direct mapping between the utilization and
 423          * the performance levels depends on the frequency invariance.
 424          */
 425         if (!arch_scale_freq_invariant()) {
 426                 sugov_update_single_freq(hook, time, flags);
 427                 return;
 428         }
 429
 430         max_cap = arch_scale_cpu_capacity(sg_cpu->cpu);
 431
 432         if (!sugov_update_single_common(sg_cpu, time, max_cap, flags))
 433                 return;
 434
 435         /*
 436          * Do not reduce the target performance level if the CPU has not been
 437          * idle recently, as the reduction is likely to be premature then.
 438          *
 439          * Except when the rq is capped by uclamp_max.
 440          */
 441         if (!uclamp_rq_is_capped(cpu_rq(sg_cpu->cpu)) &&
 442             sugov_cpu_is_busy(sg_cpu) && sg_cpu->util < prev_util)
 443                 sg_cpu->util = prev_util;
 444
 445         cpufreq_driver_adjust_perf(sg_cpu->cpu, sg_cpu->bw_min,
 446                                    sg_cpu->util, max_cap);
 447
 448         sg_cpu->sg_policy->last_freq_update_time = time;
 449 }
 450
 451 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
 452 {
 453         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 454         struct cpufreq_policy *policy = sg_policy->policy;
 455         unsigned long util = 0, max_cap;
 456         unsigned int j;
 457
 458         max_cap = arch_scale_cpu_capacity(sg_cpu->cpu);
 459
 460         for_each_cpu(j, policy->cpus) {
 461                 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
 462                 unsigned long boost;
 463
 464                 boost = sugov_iowait_apply(j_sg_cpu, time, max_cap);
 465                 sugov_get_util(j_sg_cpu, boost);
 466
 467                 util = max(j_sg_cpu->util, util);
 468         }
 469
 470         return get_next_freq(sg_policy, util, max_cap);
 471 }
 472
 473 static void
 474 sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
 475 {
 476         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 477         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 478         unsigned int next_f;
 479
 480         raw_spin_lock(&sg_policy->update_lock);
 481
 482         sugov_iowait_boost(sg_cpu, time, flags);
 483         sg_cpu->last_update = time;
 484
 485         ignore_dl_rate_limit(sg_cpu);
 486
 487         if (sugov_should_update_freq(sg_policy, time)) {
 488                 next_f = sugov_next_freq_shared(sg_cpu, time);
 489
 490                 if (!sugov_update_next_freq(sg_policy, time, next_f))
 491                         goto unlock;
 492
 493                 if (sg_policy->policy->fast_switch_enabled)
 494                         cpufreq_driver_fast_switch(sg_policy->policy, next_f);
 495                 else
 496                         sugov_deferred_update(sg_policy);
 497         }
 498 unlock:
 499         raw_spin_unlock(&sg_policy->update_lock);
 500 }
 501
 502 static void sugov_work(struct kthread_work *work)
 503 {
 504         struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
 505         unsigned int freq;
 506         unsigned long flags;
 507
 508         /*
 509          * Hold sg_policy->update_lock shortly to handle the case where:
 510          * in case sg_policy->next_freq is read here, and then updated by
 511          * sugov_deferred_update() just before work_in_progress is set to false
 512          * here, we may miss queueing the new update.
 513          *
 514          * Note: If a work was queued after the update_lock is released,
 515          * sugov_work() will just be called again by kthread_work code; and the
 516          * request will be proceed before the sugov thread sleeps.
 517          */
 518         raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
 519         freq = sg_policy->next_freq;
 520         sg_policy->work_in_progress = false;
 521         raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
 522
 523         mutex_lock(&sg_policy->work_lock);
 524         __cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
 525         mutex_unlock(&sg_policy->work_lock);
 526 }
 527
 528 static void sugov_irq_work(struct irq_work *irq_work)
 529 {
 530         struct sugov_policy *sg_policy;
 531
 532         sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
 533
 534         kthread_queue_work(&sg_policy->worker, &sg_policy->work);
 535 }
 536
 537 /************************** sysfs interface ************************/
 538
 539 static struct sugov_tunables *global_tunables;
 540 static DEFINE_MUTEX(global_tunables_lock);
 541
 542 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
 543 {
 544         return container_of(attr_set, struct sugov_tunables, attr_set);
 545 }
 546
 547 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
 548 {
 549         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 550
 551         return sprintf(buf, "%u\n", tunables->rate_limit_us);
 552 }
 553
 554 static ssize_t
 555 rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
 556 {
 557         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 558         struct sugov_policy *sg_policy;
 559         unsigned int rate_limit_us;
 560
 561         if (kstrtouint(buf, 10, &rate_limit_us))
 562                 return -EINVAL;
 563
 564         tunables->rate_limit_us = rate_limit_us;
 565
 566         list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
 567                 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
 568
 569         return count;
 570 }
 571
 572 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
 573
 574 static struct attribute *sugov_attrs[] = {
 575         &rate_limit_us.attr,
 576         NULL
 577 };
 578 ATTRIBUTE_GROUPS(sugov);
 579
 580 static void sugov_tunables_free(struct kobject *kobj)
 581 {
 582         struct gov_attr_set *attr_set = to_gov_attr_set(kobj);
 583
 584         kfree(to_sugov_tunables(attr_set));
 585 }
 586
 587 static const struct kobj_type sugov_tunables_ktype = {
 588         .default_groups = sugov_groups,
 589         .sysfs_ops = &governor_sysfs_ops,
 590         .release = &sugov_tunables_free,
 591 };
 592
 593 /********************** cpufreq governor interface *********************/
 594
 595 #ifdef CONFIG_ENERGY_MODEL
 596 static void rebuild_sd_workfn(struct work_struct *work)
 597 {
 598         rebuild_sched_domains_energy();
 599 }
 600
 601 static DECLARE_WORK(rebuild_sd_work, rebuild_sd_workfn);
 602
 603 /*
 604  * EAS shouldn't be attempted without sugov, so rebuild the sched_domains
 605  * on governor changes to make sure the scheduler knows about it.
 606  */
 607 static void sugov_eas_rebuild_sd(void)
 608 {
 609         /*
 610          * When called from the cpufreq_register_driver() path, the
 611          * cpu_hotplug_lock is already held, so use a work item to
 612          * avoid nested locking in rebuild_sched_domains().
 613          */
 614         schedule_work(&rebuild_sd_work);
 615 }
 616 #else
 617 static inline void sugov_eas_rebuild_sd(void) { };
 618 #endif
 619
 620 struct cpufreq_governor schedutil_gov;
 621
 622 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
 623 {
 624         struct sugov_policy *sg_policy;
 625
 626         sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
 627         if (!sg_policy)
 628                 return NULL;
 629
 630         sg_policy->policy = policy;
 631         raw_spin_lock_init(&sg_policy->update_lock);
 632         return sg_policy;
 633 }
 634
 635 static void sugov_policy_free(struct sugov_policy *sg_policy)
 636 {
 637         kfree(sg_policy);
 638 }
 639
 640 static int sugov_kthread_create(struct sugov_policy *sg_policy)
 641 {
 642         struct task_struct *thread;
 643         struct sched_attr attr = {
 644                 .size           = sizeof(struct sched_attr),
 645                 .sched_policy   = SCHED_DEADLINE,
 646                 .sched_flags    = SCHED_FLAG_SUGOV,
 647                 .sched_nice     = 0,
 648                 .sched_priority = 0,
 649                 /*
 650                  * Fake (unused) bandwidth; workaround to "fix"
 651                  * priority inheritance.
 652                  */
 653                 .sched_runtime  =  1000000,
 654                 .sched_deadline = 10000000,
 655                 .sched_period   = 10000000,
 656         };
 657         struct cpufreq_policy *policy = sg_policy->policy;
 658         int ret;
 659
 660         /* kthread only required for slow path */
 661         if (policy->fast_switch_enabled)
 662                 return 0;
 663
 664         kthread_init_work(&sg_policy->work, sugov_work);
 665         kthread_init_worker(&sg_policy->worker);
 666         thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
 667                                 "sugov:%d",
 668                                 cpumask_first(policy->related_cpus));
 669         if (IS_ERR(thread)) {
 670                 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
 671                 return PTR_ERR(thread);
 672         }
 673
 674         ret = sched_setattr_nocheck(thread, &attr);
 675         if (ret) {
 676                 kthread_stop(thread);
 677                 pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
 678                 return ret;
 679         }
 680
 681         sg_policy->thread = thread;
 682         kthread_bind_mask(thread, policy->related_cpus);
 683         init_irq_work(&sg_policy->irq_work, sugov_irq_work);
 684         mutex_init(&sg_policy->work_lock);
 685
 686         wake_up_process(thread);
 687
 688         return 0;
 689 }
 690
 691 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
 692 {
 693         /* kthread only required for slow path */
 694         if (sg_policy->policy->fast_switch_enabled)
 695                 return;
 696
 697         kthread_flush_worker(&sg_policy->worker);
 698         kthread_stop(sg_policy->thread);
 699         mutex_destroy(&sg_policy->work_lock);
 700 }
 701
 702 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
 703 {
 704         struct sugov_tunables *tunables;
 705
 706         tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
 707         if (tunables) {
 708                 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
 709                 if (!have_governor_per_policy())
 710                         global_tunables = tunables;
 711         }
 712         return tunables;
 713 }
 714
 715 static void sugov_clear_global_tunables(void)
 716 {
 717         if (!have_governor_per_policy())
 718                 global_tunables = NULL;
 719 }
 720
 721 static int sugov_init(struct cpufreq_policy *policy)
 722 {
 723         struct sugov_policy *sg_policy;
 724         struct sugov_tunables *tunables;
 725         int ret = 0;
 726
 727         /* State should be equivalent to EXIT */
 728         if (policy->governor_data)
 729                 return -EBUSY;
 730
 731         cpufreq_enable_fast_switch(policy);
 732
 733         sg_policy = sugov_policy_alloc(policy);
 734         if (!sg_policy) {
 735                 ret = -ENOMEM;
 736                 goto disable_fast_switch;
 737         }
 738
 739         ret = sugov_kthread_create(sg_policy);
 740         if (ret)
 741                 goto free_sg_policy;
 742
 743         mutex_lock(&global_tunables_lock);
 744
 745         if (global_tunables) {
 746                 if (WARN_ON(have_governor_per_policy())) {
 747                         ret = -EINVAL;
 748                         goto stop_kthread;
 749                 }
 750                 policy->governor_data = sg_policy;
 751                 sg_policy->tunables = global_tunables;
 752
 753                 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
 754                 goto out;
 755         }
 756
 757         tunables = sugov_tunables_alloc(sg_policy);
 758         if (!tunables) {
 759                 ret = -ENOMEM;
 760                 goto stop_kthread;
 761         }
 762
 763         tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
 764
 765         policy->governor_data = sg_policy;
 766         sg_policy->tunables = tunables;
 767
 768         ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
 769                                    get_governor_parent_kobj(policy), "%s",
 770                                    schedutil_gov.name);
 771         if (ret)
 772                 goto fail;
 773
 774         sugov_eas_rebuild_sd();
 775
 776 out:
 777         mutex_unlock(&global_tunables_lock);
 778         return 0;
 779
 780 fail:
 781         kobject_put(&tunables->attr_set.kobj);
 782         policy->governor_data = NULL;
 783         sugov_clear_global_tunables();
 784
 785 stop_kthread:
 786         sugov_kthread_stop(sg_policy);
 787         mutex_unlock(&global_tunables_lock);
 788
 789 free_sg_policy:
 790         sugov_policy_free(sg_policy);
 791
 792 disable_fast_switch:
 793         cpufreq_disable_fast_switch(policy);
 794
 795         pr_err("initialization failed (error %d)\n", ret);
 796         return ret;
 797 }
 798
 799 static void sugov_exit(struct cpufreq_policy *policy)
 800 {
 801         struct sugov_policy *sg_policy = policy->governor_data;
 802         struct sugov_tunables *tunables = sg_policy->tunables;
 803         unsigned int count;
 804
 805         mutex_lock(&global_tunables_lock);
 806
 807         count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
 808         policy->governor_data = NULL;
 809         if (!count)
 810                 sugov_clear_global_tunables();
 811
 812         mutex_unlock(&global_tunables_lock);
 813
 814         sugov_kthread_stop(sg_policy);
 815         sugov_policy_free(sg_policy);
 816         cpufreq_disable_fast_switch(policy);
 817
 818         sugov_eas_rebuild_sd();
 819 }
 820
 821 static int sugov_start(struct cpufreq_policy *policy)
 822 {
 823         struct sugov_policy *sg_policy = policy->governor_data;
 824         void (*uu)(struct update_util_data *data, u64 time, unsigned int flags);
 825         unsigned int cpu;
 826
 827         sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
 828         sg_policy->last_freq_update_time        = 0;
 829         sg_policy->next_freq                    = 0;
 830         sg_policy->work_in_progress             = false;
 831         sg_policy->limits_changed               = false;
 832         sg_policy->cached_raw_freq              = 0;
 833
 834         sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
 835
 836         if (policy_is_shared(policy))
 837                 uu = sugov_update_shared;
 838         else if (policy->fast_switch_enabled && cpufreq_driver_has_adjust_perf())
 839                 uu = sugov_update_single_perf;
 840         else
 841                 uu = sugov_update_single_freq;
 842
 843         for_each_cpu(cpu, policy->cpus) {
 844                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 845
 846                 memset(sg_cpu, 0, sizeof(*sg_cpu));
 847                 sg_cpu->cpu = cpu;
 848                 sg_cpu->sg_policy = sg_policy;
 849                 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util, uu);
 850         }
 851         return 0;
 852 }
 853
 854 static void sugov_stop(struct cpufreq_policy *policy)
 855 {
 856         struct sugov_policy *sg_policy = policy->governor_data;
 857         unsigned int cpu;
 858
 859         for_each_cpu(cpu, policy->cpus)
 860                 cpufreq_remove_update_util_hook(cpu);
 861
 862         synchronize_rcu();
 863
 864         if (!policy->fast_switch_enabled) {
 865                 irq_work_sync(&sg_policy->irq_work);
 866                 kthread_cancel_work_sync(&sg_policy->work);
 867         }
 868 }
 869
 870 static void sugov_limits(struct cpufreq_policy *policy)
 871 {
 872         struct sugov_policy *sg_policy = policy->governor_data;
 873
 874         if (!policy->fast_switch_enabled) {
 875                 mutex_lock(&sg_policy->work_lock);
 876                 cpufreq_policy_apply_limits(policy);
 877                 mutex_unlock(&sg_policy->work_lock);
 878         }
 879
 880         sg_policy->limits_changed = true;
 881 }
 882
 883 struct cpufreq_governor schedutil_gov = {
 884         .name                   = "schedutil",
 885         .owner                  = THIS_MODULE,
 886         .flags                  = CPUFREQ_GOV_DYNAMIC_SWITCHING,
 887         .init                   = sugov_init,
 888         .exit                   = sugov_exit,
 889         .start                  = sugov_start,
 890         .stop                   = sugov_stop,
 891         .limits                 = sugov_limits,
 892 };
 893
 894 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
 895 struct cpufreq_governor *cpufreq_default_governor(void)
 896 {
 897         return &schedutil_gov;
 898 }
 899 #endif
 900
 901 cpufreq_governor_init(schedutil_gov);