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         /*
 137          * Apply a 25% margin so that we select a higher frequency than
 138          * the current one before the CPU is fully busy:
 139          */
 140         return policy->cur + (policy->cur >> 2);
 141 }
 142
 143 /**
 144  * get_next_freq - Compute a new frequency for a given cpufreq policy.
 145  * @sg_policy: schedutil policy object to compute the new frequency for.
 146  * @util: Current CPU utilization.
 147  * @max: CPU capacity.
 148  *
 149  * If the utilization is frequency-invariant, choose the new frequency to be
 150  * proportional to it, that is
 151  *
 152  * next_freq = C * max_freq * util / max
 153  *
 154  * Otherwise, approximate the would-be frequency-invariant utilization by
 155  * util_raw * (curr_freq / max_freq) which leads to
 156  *
 157  * next_freq = C * curr_freq * util_raw / max
 158  *
 159  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
 160  *
 161  * The lowest driver-supported frequency which is equal or greater than the raw
 162  * next_freq (as calculated above) is returned, subject to policy min/max and
 163  * cpufreq driver limitations.
 164  */
 165 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
 166                                   unsigned long util, unsigned long max)
 167 {
 168         struct cpufreq_policy *policy = sg_policy->policy;
 169         unsigned int freq;
 170
 171         freq = get_capacity_ref_freq(policy);
 172         freq = map_util_freq(util, freq, max);
 173
 174         if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
 175                 return sg_policy->next_freq;
 176
 177         sg_policy->cached_raw_freq = freq;
 178         return cpufreq_driver_resolve_freq(policy, freq);
 179 }
 180
 181 unsigned long sugov_effective_cpu_perf(int cpu, unsigned long actual,
 182                                  unsigned long min,
 183                                  unsigned long max)
 184 {
 185         /* Add dvfs headroom to actual utilization */
 186         actual = map_util_perf(actual);
 187         /* Actually we don't need to target the max performance */
 188         if (actual < max)
 189                 max = actual;
 190
 191         /*
 192          * Ensure at least minimum performance while providing more compute
 193          * capacity when possible.
 194          */
 195         return max(min, max);
 196 }
 197
 198 static void sugov_get_util(struct sugov_cpu *sg_cpu, unsigned long boost)
 199 {
 200         unsigned long min, max, util = cpu_util_cfs_boost(sg_cpu->cpu);
 201
 202         util = effective_cpu_util(sg_cpu->cpu, util, &min, &max);
 203         util = max(util, boost);
 204         sg_cpu->bw_min = min;
 205         sg_cpu->util = sugov_effective_cpu_perf(sg_cpu->cpu, util, min, max);
 206 }
 207
 208 /**
 209  * sugov_iowait_reset() - Reset the IO boost status of a CPU.
 210  * @sg_cpu: the sugov data for the CPU to boost
 211  * @time: the update time from the caller
 212  * @set_iowait_boost: true if an IO boost has been requested
 213  *
 214  * The IO wait boost of a task is disabled after a tick since the last update
 215  * of a CPU. If a new IO wait boost is requested after more then a tick, then
 216  * we enable the boost starting from IOWAIT_BOOST_MIN, which improves energy
 217  * efficiency by ignoring sporadic wakeups from IO.
 218  */
 219 static bool sugov_iowait_reset(struct sugov_cpu *sg_cpu, u64 time,
 220                                bool set_iowait_boost)
 221 {
 222         s64 delta_ns = time - sg_cpu->last_update;
 223
 224         /* Reset boost only if a tick has elapsed since last request */
 225         if (delta_ns <= TICK_NSEC)
 226                 return false;
 227
 228         sg_cpu->iowait_boost = set_iowait_boost ? IOWAIT_BOOST_MIN : 0;
 229         sg_cpu->iowait_boost_pending = set_iowait_boost;
 230
 231         return true;
 232 }
 233
 234 /**
 235  * sugov_iowait_boost() - Updates the IO boost status of a CPU.
 236  * @sg_cpu: the sugov data for the CPU to boost
 237  * @time: the update time from the caller
 238  * @flags: SCHED_CPUFREQ_IOWAIT if the task is waking up after an IO wait
 239  *
 240  * Each time a task wakes up after an IO operation, the CPU utilization can be
 241  * boosted to a certain utilization which doubles at each "frequent and
 242  * successive" wakeup from IO, ranging from IOWAIT_BOOST_MIN to the utilization
 243  * of the maximum OPP.
 244  *
 245  * To keep doubling, an IO boost has to be requested at least once per tick,
 246  * otherwise we restart from the utilization of the minimum OPP.
 247  */
 248 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
 249                                unsigned int flags)
 250 {
 251         bool set_iowait_boost = flags & SCHED_CPUFREQ_IOWAIT;
 252
 253         /* Reset boost if the CPU appears to have been idle enough */
 254         if (sg_cpu->iowait_boost &&
 255             sugov_iowait_reset(sg_cpu, time, set_iowait_boost))
 256                 return;
 257
 258         /* Boost only tasks waking up after IO */
 259         if (!set_iowait_boost)
 260                 return;
 261
 262         /* Ensure boost doubles only one time at each request */
 263         if (sg_cpu->iowait_boost_pending)
 264                 return;
 265         sg_cpu->iowait_boost_pending = true;
 266
 267         /* Double the boost at each request */
 268         if (sg_cpu->iowait_boost) {
 269                 sg_cpu->iowait_boost =
 270                         min_t(unsigned int, sg_cpu->iowait_boost << 1, SCHED_CAPACITY_SCALE);
 271                 return;
 272         }
 273
 274         /* First wakeup after IO: start with minimum boost */
 275         sg_cpu->iowait_boost = IOWAIT_BOOST_MIN;
 276 }
 277
 278 /**
 279  * sugov_iowait_apply() - Apply the IO boost to a CPU.
 280  * @sg_cpu: the sugov data for the cpu to boost
 281  * @time: the update time from the caller
 282  * @max_cap: the max CPU capacity
 283  *
 284  * A CPU running a task which woken up after an IO operation can have its
 285  * utilization boosted to speed up the completion of those IO operations.
 286  * The IO boost value is increased each time a task wakes up from IO, in
 287  * sugov_iowait_apply(), and it's instead decreased by this function,
 288  * each time an increase has not been requested (!iowait_boost_pending).
 289  *
 290  * A CPU which also appears to have been idle for at least one tick has also
 291  * its IO boost utilization reset.
 292  *
 293  * This mechanism is designed to boost high frequently IO waiting tasks, while
 294  * being more conservative on tasks which does sporadic IO operations.
 295  */
 296 static unsigned long sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
 297                                unsigned long max_cap)
 298 {
 299         /* No boost currently required */
 300         if (!sg_cpu->iowait_boost)
 301                 return 0;
 302
 303         /* Reset boost if the CPU appears to have been idle enough */
 304         if (sugov_iowait_reset(sg_cpu, time, false))
 305                 return 0;
 306
 307         if (!sg_cpu->iowait_boost_pending) {
 308                 /*
 309                  * No boost pending; reduce the boost value.
 310                  */
 311                 sg_cpu->iowait_boost >>= 1;
 312                 if (sg_cpu->iowait_boost < IOWAIT_BOOST_MIN) {
 313                         sg_cpu->iowait_boost = 0;
 314                         return 0;
 315                 }
 316         }
 317
 318         sg_cpu->iowait_boost_pending = false;
 319
 320         /*
 321          * sg_cpu->util is already in capacity scale; convert iowait_boost
 322          * into the same scale so we can compare.
 323          */
 324         return (sg_cpu->iowait_boost * max_cap) >> SCHED_CAPACITY_SHIFT;
 325 }
 326
 327 #ifdef CONFIG_NO_HZ_COMMON
 328 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
 329 {
 330         unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
 331         bool ret = idle_calls == sg_cpu->saved_idle_calls;
 332
 333         sg_cpu->saved_idle_calls = idle_calls;
 334         return ret;
 335 }
 336 #else
 337 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
 338 #endif /* CONFIG_NO_HZ_COMMON */
 339
 340 /*
 341  * Make sugov_should_update_freq() ignore the rate limit when DL
 342  * has increased the utilization.
 343  */
 344 static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu)
 345 {
 346         if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_min)
 347                 sg_cpu->sg_policy->limits_changed = true;
 348 }
 349
 350 static inline bool sugov_update_single_common(struct sugov_cpu *sg_cpu,
 351                                               u64 time, unsigned long max_cap,
 352                                               unsigned int flags)
 353 {
 354         unsigned long boost;
 355
 356         sugov_iowait_boost(sg_cpu, time, flags);
 357         sg_cpu->last_update = time;
 358
 359         ignore_dl_rate_limit(sg_cpu);
 360
 361         if (!sugov_should_update_freq(sg_cpu->sg_policy, time))
 362                 return false;
 363
 364         boost = sugov_iowait_apply(sg_cpu, time, max_cap);
 365         sugov_get_util(sg_cpu, boost);
 366
 367         return true;
 368 }
 369
 370 static void sugov_update_single_freq(struct update_util_data *hook, u64 time,
 371                                      unsigned int flags)
 372 {
 373         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 374         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 375         unsigned int cached_freq = sg_policy->cached_raw_freq;
 376         unsigned long max_cap;
 377         unsigned int next_f;
 378
 379         max_cap = arch_scale_cpu_capacity(sg_cpu->cpu);
 380
 381         if (!sugov_update_single_common(sg_cpu, time, max_cap, flags))
 382                 return;
 383
 384         next_f = get_next_freq(sg_policy, sg_cpu->util, max_cap);
 385         /*
 386          * Do not reduce the frequency if the CPU has not been idle
 387          * recently, as the reduction is likely to be premature then.
 388          *
 389          * Except when the rq is capped by uclamp_max.
 390          */
 391         if (!uclamp_rq_is_capped(cpu_rq(sg_cpu->cpu)) &&
 392             sugov_cpu_is_busy(sg_cpu) && next_f < sg_policy->next_freq &&
 393             !sg_policy->need_freq_update) {
 394                 next_f = sg_policy->next_freq;
 395
 396                 /* Restore cached freq as next_freq has changed */
 397                 sg_policy->cached_raw_freq = cached_freq;
 398         }
 399
 400         if (!sugov_update_next_freq(sg_policy, time, next_f))
 401                 return;
 402
 403         /*
 404          * This code runs under rq->lock for the target CPU, so it won't run
 405          * concurrently on two different CPUs for the same target and it is not
 406          * necessary to acquire the lock in the fast switch case.
 407          */
 408         if (sg_policy->policy->fast_switch_enabled) {
 409                 cpufreq_driver_fast_switch(sg_policy->policy, next_f);
 410         } else {
 411                 raw_spin_lock(&sg_policy->update_lock);
 412                 sugov_deferred_update(sg_policy);
 413                 raw_spin_unlock(&sg_policy->update_lock);
 414         }
 415 }
 416
 417 static void sugov_update_single_perf(struct update_util_data *hook, u64 time,
 418                                      unsigned int flags)
 419 {
 420         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 421         unsigned long prev_util = sg_cpu->util;
 422         unsigned long max_cap;
 423
 424         /*
 425          * Fall back to the "frequency" path if frequency invariance is not
 426          * supported, because the direct mapping between the utilization and
 427          * the performance levels depends on the frequency invariance.
 428          */
 429         if (!arch_scale_freq_invariant()) {
 430                 sugov_update_single_freq(hook, time, flags);
 431                 return;
 432         }
 433
 434         max_cap = arch_scale_cpu_capacity(sg_cpu->cpu);
 435
 436         if (!sugov_update_single_common(sg_cpu, time, max_cap, flags))
 437                 return;
 438
 439         /*
 440          * Do not reduce the target performance level if the CPU has not been
 441          * idle recently, as the reduction is likely to be premature then.
 442          *
 443          * Except when the rq is capped by uclamp_max.
 444          */
 445         if (!uclamp_rq_is_capped(cpu_rq(sg_cpu->cpu)) &&
 446             sugov_cpu_is_busy(sg_cpu) && sg_cpu->util < prev_util)
 447                 sg_cpu->util = prev_util;
 448
 449         cpufreq_driver_adjust_perf(sg_cpu->cpu, sg_cpu->bw_min,
 450                                    sg_cpu->util, max_cap);
 451
 452         sg_cpu->sg_policy->last_freq_update_time = time;
 453 }
 454
 455 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
 456 {
 457         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 458         struct cpufreq_policy *policy = sg_policy->policy;
 459         unsigned long util = 0, max_cap;
 460         unsigned int j;
 461
 462         max_cap = arch_scale_cpu_capacity(sg_cpu->cpu);
 463
 464         for_each_cpu(j, policy->cpus) {
 465                 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
 466                 unsigned long boost;
 467
 468                 boost = sugov_iowait_apply(j_sg_cpu, time, max_cap);
 469                 sugov_get_util(j_sg_cpu, boost);
 470
 471                 util = max(j_sg_cpu->util, util);
 472         }
 473
 474         return get_next_freq(sg_policy, util, max_cap);
 475 }
 476
 477 static void
 478 sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
 479 {
 480         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 481         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 482         unsigned int next_f;
 483
 484         raw_spin_lock(&sg_policy->update_lock);
 485
 486         sugov_iowait_boost(sg_cpu, time, flags);
 487         sg_cpu->last_update = time;
 488
 489         ignore_dl_rate_limit(sg_cpu);
 490
 491         if (sugov_should_update_freq(sg_policy, time)) {
 492                 next_f = sugov_next_freq_shared(sg_cpu, time);
 493
 494                 if (!sugov_update_next_freq(sg_policy, time, next_f))
 495                         goto unlock;
 496
 497                 if (sg_policy->policy->fast_switch_enabled)
 498                         cpufreq_driver_fast_switch(sg_policy->policy, next_f);
 499                 else
 500                         sugov_deferred_update(sg_policy);
 501         }
 502 unlock:
 503         raw_spin_unlock(&sg_policy->update_lock);
 504 }
 505
 506 static void sugov_work(struct kthread_work *work)
 507 {
 508         struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
 509         unsigned int freq;
 510         unsigned long flags;
 511
 512         /*
 513          * Hold sg_policy->update_lock shortly to handle the case where:
 514          * in case sg_policy->next_freq is read here, and then updated by
 515          * sugov_deferred_update() just before work_in_progress is set to false
 516          * here, we may miss queueing the new update.
 517          *
 518          * Note: If a work was queued after the update_lock is released,
 519          * sugov_work() will just be called again by kthread_work code; and the
 520          * request will be proceed before the sugov thread sleeps.
 521          */
 522         raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
 523         freq = sg_policy->next_freq;
 524         sg_policy->work_in_progress = false;
 525         raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
 526
 527         mutex_lock(&sg_policy->work_lock);
 528         __cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
 529         mutex_unlock(&sg_policy->work_lock);
 530 }
 531
 532 static void sugov_irq_work(struct irq_work *irq_work)
 533 {
 534         struct sugov_policy *sg_policy;
 535
 536         sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
 537
 538         kthread_queue_work(&sg_policy->worker, &sg_policy->work);
 539 }
 540
 541 /************************** sysfs interface ************************/
 542
 543 static struct sugov_tunables *global_tunables;
 544 static DEFINE_MUTEX(global_tunables_lock);
 545
 546 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
 547 {
 548         return container_of(attr_set, struct sugov_tunables, attr_set);
 549 }
 550
 551 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
 552 {
 553         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 554
 555         return sprintf(buf, "%u\n", tunables->rate_limit_us);
 556 }
 557
 558 static ssize_t
 559 rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
 560 {
 561         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 562         struct sugov_policy *sg_policy;
 563         unsigned int rate_limit_us;
 564
 565         if (kstrtouint(buf, 10, &rate_limit_us))
 566                 return -EINVAL;
 567
 568         tunables->rate_limit_us = rate_limit_us;
 569
 570         list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
 571                 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
 572
 573         return count;
 574 }
 575
 576 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
 577
 578 static struct attribute *sugov_attrs[] = {
 579         &rate_limit_us.attr,
 580         NULL
 581 };
 582 ATTRIBUTE_GROUPS(sugov);
 583
 584 static void sugov_tunables_free(struct kobject *kobj)
 585 {
 586         struct gov_attr_set *attr_set = to_gov_attr_set(kobj);
 587
 588         kfree(to_sugov_tunables(attr_set));
 589 }
 590
 591 static const struct kobj_type sugov_tunables_ktype = {
 592         .default_groups = sugov_groups,
 593         .sysfs_ops = &governor_sysfs_ops,
 594         .release = &sugov_tunables_free,
 595 };
 596
 597 /********************** cpufreq governor interface *********************/
 598
 599 #ifdef CONFIG_ENERGY_MODEL
 600 static void rebuild_sd_workfn(struct work_struct *work)
 601 {
 602         rebuild_sched_domains_energy();
 603 }
 604
 605 static DECLARE_WORK(rebuild_sd_work, rebuild_sd_workfn);
 606
 607 /*
 608  * EAS shouldn't be attempted without sugov, so rebuild the sched_domains
 609  * on governor changes to make sure the scheduler knows about it.
 610  */
 611 static void sugov_eas_rebuild_sd(void)
 612 {
 613         /*
 614          * When called from the cpufreq_register_driver() path, the
 615          * cpu_hotplug_lock is already held, so use a work item to
 616          * avoid nested locking in rebuild_sched_domains().
 617          */
 618         schedule_work(&rebuild_sd_work);
 619 }
 620 #else
 621 static inline void sugov_eas_rebuild_sd(void) { };
 622 #endif
 623
 624 struct cpufreq_governor schedutil_gov;
 625
 626 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
 627 {
 628         struct sugov_policy *sg_policy;
 629
 630         sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
 631         if (!sg_policy)
 632                 return NULL;
 633
 634         sg_policy->policy = policy;
 635         raw_spin_lock_init(&sg_policy->update_lock);
 636         return sg_policy;
 637 }
 638
 639 static void sugov_policy_free(struct sugov_policy *sg_policy)
 640 {
 641         kfree(sg_policy);
 642 }
 643
 644 static int sugov_kthread_create(struct sugov_policy *sg_policy)
 645 {
 646         struct task_struct *thread;
 647         struct sched_attr attr = {
 648                 .size           = sizeof(struct sched_attr),
 649                 .sched_policy   = SCHED_DEADLINE,
 650                 .sched_flags    = SCHED_FLAG_SUGOV,
 651                 .sched_nice     = 0,
 652                 .sched_priority = 0,
 653                 /*
 654                  * Fake (unused) bandwidth; workaround to "fix"
 655                  * priority inheritance.
 656                  */
 657                 .sched_runtime  =  1000000,
 658                 .sched_deadline = 10000000,
 659                 .sched_period   = 10000000,
 660         };
 661         struct cpufreq_policy *policy = sg_policy->policy;
 662         int ret;
 663
 664         /* kthread only required for slow path */
 665         if (policy->fast_switch_enabled)
 666                 return 0;
 667
 668         kthread_init_work(&sg_policy->work, sugov_work);
 669         kthread_init_worker(&sg_policy->worker);
 670         thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
 671                                 "sugov:%d",
 672                                 cpumask_first(policy->related_cpus));
 673         if (IS_ERR(thread)) {
 674                 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
 675                 return PTR_ERR(thread);
 676         }
 677
 678         ret = sched_setattr_nocheck(thread, &attr);
 679         if (ret) {
 680                 kthread_stop(thread);
 681                 pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
 682                 return ret;
 683         }
 684
 685         sg_policy->thread = thread;
 686         kthread_bind_mask(thread, policy->related_cpus);
 687         init_irq_work(&sg_policy->irq_work, sugov_irq_work);
 688         mutex_init(&sg_policy->work_lock);
 689
 690         wake_up_process(thread);
 691
 692         return 0;
 693 }
 694
 695 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
 696 {
 697         /* kthread only required for slow path */
 698         if (sg_policy->policy->fast_switch_enabled)
 699                 return;
 700
 701         kthread_flush_worker(&sg_policy->worker);
 702         kthread_stop(sg_policy->thread);
 703         mutex_destroy(&sg_policy->work_lock);
 704 }
 705
 706 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
 707 {
 708         struct sugov_tunables *tunables;
 709
 710         tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
 711         if (tunables) {
 712                 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
 713                 if (!have_governor_per_policy())
 714                         global_tunables = tunables;
 715         }
 716         return tunables;
 717 }
 718
 719 static void sugov_clear_global_tunables(void)
 720 {
 721         if (!have_governor_per_policy())
 722                 global_tunables = NULL;
 723 }
 724
 725 static int sugov_init(struct cpufreq_policy *policy)
 726 {
 727         struct sugov_policy *sg_policy;
 728         struct sugov_tunables *tunables;
 729         int ret = 0;
 730
 731         /* State should be equivalent to EXIT */
 732         if (policy->governor_data)
 733                 return -EBUSY;
 734
 735         cpufreq_enable_fast_switch(policy);
 736
 737         sg_policy = sugov_policy_alloc(policy);
 738         if (!sg_policy) {
 739                 ret = -ENOMEM;
 740                 goto disable_fast_switch;
 741         }
 742
 743         ret = sugov_kthread_create(sg_policy);
 744         if (ret)
 745                 goto free_sg_policy;
 746
 747         mutex_lock(&global_tunables_lock);
 748
 749         if (global_tunables) {
 750                 if (WARN_ON(have_governor_per_policy())) {
 751                         ret = -EINVAL;
 752                         goto stop_kthread;
 753                 }
 754                 policy->governor_data = sg_policy;
 755                 sg_policy->tunables = global_tunables;
 756
 757                 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
 758                 goto out;
 759         }
 760
 761         tunables = sugov_tunables_alloc(sg_policy);
 762         if (!tunables) {
 763                 ret = -ENOMEM;
 764                 goto stop_kthread;
 765         }
 766
 767         tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
 768
 769         policy->governor_data = sg_policy;
 770         sg_policy->tunables = tunables;
 771
 772         ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
 773                                    get_governor_parent_kobj(policy), "%s",
 774                                    schedutil_gov.name);
 775         if (ret)
 776                 goto fail;
 777
 778         sugov_eas_rebuild_sd();
 779
 780 out:
 781         mutex_unlock(&global_tunables_lock);
 782         return 0;
 783
 784 fail:
 785         kobject_put(&tunables->attr_set.kobj);
 786         policy->governor_data = NULL;
 787         sugov_clear_global_tunables();
 788
 789 stop_kthread:
 790         sugov_kthread_stop(sg_policy);
 791         mutex_unlock(&global_tunables_lock);
 792
 793 free_sg_policy:
 794         sugov_policy_free(sg_policy);
 795
 796 disable_fast_switch:
 797         cpufreq_disable_fast_switch(policy);
 798
 799         pr_err("initialization failed (error %d)\n", ret);
 800         return ret;
 801 }
 802
 803 static void sugov_exit(struct cpufreq_policy *policy)
 804 {
 805         struct sugov_policy *sg_policy = policy->governor_data;
 806         struct sugov_tunables *tunables = sg_policy->tunables;
 807         unsigned int count;
 808
 809         mutex_lock(&global_tunables_lock);
 810
 811         count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
 812         policy->governor_data = NULL;
 813         if (!count)
 814                 sugov_clear_global_tunables();
 815
 816         mutex_unlock(&global_tunables_lock);
 817
 818         sugov_kthread_stop(sg_policy);
 819         sugov_policy_free(sg_policy);
 820         cpufreq_disable_fast_switch(policy);
 821
 822         sugov_eas_rebuild_sd();
 823 }
 824
 825 static int sugov_start(struct cpufreq_policy *policy)
 826 {
 827         struct sugov_policy *sg_policy = policy->governor_data;
 828         void (*uu)(struct update_util_data *data, u64 time, unsigned int flags);
 829         unsigned int cpu;
 830
 831         sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
 832         sg_policy->last_freq_update_time        = 0;
 833         sg_policy->next_freq                    = 0;
 834         sg_policy->work_in_progress             = false;
 835         sg_policy->limits_changed               = false;
 836         sg_policy->cached_raw_freq              = 0;
 837
 838         sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
 839
 840         if (policy_is_shared(policy))
 841                 uu = sugov_update_shared;
 842         else if (policy->fast_switch_enabled && cpufreq_driver_has_adjust_perf())
 843                 uu = sugov_update_single_perf;
 844         else
 845                 uu = sugov_update_single_freq;
 846
 847         for_each_cpu(cpu, policy->cpus) {
 848                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 849
 850                 memset(sg_cpu, 0, sizeof(*sg_cpu));
 851                 sg_cpu->cpu = cpu;
 852                 sg_cpu->sg_policy = sg_policy;
 853                 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util, uu);
 854         }
 855         return 0;
 856 }
 857
 858 static void sugov_stop(struct cpufreq_policy *policy)
 859 {
 860         struct sugov_policy *sg_policy = policy->governor_data;
 861         unsigned int cpu;
 862
 863         for_each_cpu(cpu, policy->cpus)
 864                 cpufreq_remove_update_util_hook(cpu);
 865
 866         synchronize_rcu();
 867
 868         if (!policy->fast_switch_enabled) {
 869                 irq_work_sync(&sg_policy->irq_work);
 870                 kthread_cancel_work_sync(&sg_policy->work);
 871         }
 872 }
 873
 874 static void sugov_limits(struct cpufreq_policy *policy)
 875 {
 876         struct sugov_policy *sg_policy = policy->governor_data;
 877
 878         if (!policy->fast_switch_enabled) {
 879                 mutex_lock(&sg_policy->work_lock);
 880                 cpufreq_policy_apply_limits(policy);
 881                 mutex_unlock(&sg_policy->work_lock);
 882         }
 883
 884         sg_policy->limits_changed = true;
 885 }
 886
 887 struct cpufreq_governor schedutil_gov = {
 888         .name                   = "schedutil",
 889         .owner                  = THIS_MODULE,
 890         .flags                  = CPUFREQ_GOV_DYNAMIC_SWITCHING,
 891         .init                   = sugov_init,
 892         .exit                   = sugov_exit,
 893         .start                  = sugov_start,
 894         .stop                   = sugov_stop,
 895         .limits                 = sugov_limits,
 896 };
 897
 898 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
 899 struct cpufreq_governor *cpufreq_default_governor(void)
 900 {
 901         return &schedutil_gov;
 902 }
 903 #endif
 904
 905 cpufreq_governor_init(schedutil_gov);