kernel/sched/cpufreq_schedutil.c

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