kernel/sched/cpufreq_schedutil.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * CPUFreq governor based on scheduler-provided CPU utilization data.
   4  *
   5  * Copyright (C) 2016, Intel Corporation
   6  * Author: Rafael J. Wysocki <[email protected]>
   7  */
   8
   9 #define 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 = scx_cpuperf_target(sg_cpu->cpu);
 201
 202         if (!scx_switched_all())
 203                 util += cpu_util_cfs_boost(sg_cpu->cpu);
 204         util = effective_cpu_util(sg_cpu->cpu, util, &min, &max);
 205         util = max(util, boost);
 206         sg_cpu->bw_min = min;
 207         sg_cpu->util = sugov_effective_cpu_perf(sg_cpu->cpu, util, min, max);
 208 }
 209
 210 /**
 211  * sugov_iowait_reset() - Reset the IO boost status of a CPU.
 212  * @sg_cpu: the sugov data for the CPU to boost
 213  * @time: the update time from the caller
 214  * @set_iowait_boost: true if an IO boost has been requested
 215  *
 216  * The IO wait boost of a task is disabled after a tick since the last update
 217  * of a CPU. If a new IO wait boost is requested after more then a tick, then
 218  * we enable the boost starting from IOWAIT_BOOST_MIN, which improves energy
 219  * efficiency by ignoring sporadic wakeups from IO.
 220  */
 221 static bool sugov_iowait_reset(struct sugov_cpu *sg_cpu, u64 time,
 222                                bool set_iowait_boost)
 223 {
 224         s64 delta_ns = time - sg_cpu->last_update;
 225
 226         /* Reset boost only if a tick has elapsed since last request */
 227         if (delta_ns <= TICK_NSEC)
 228                 return false;
 229
 230         sg_cpu->iowait_boost = set_iowait_boost ? IOWAIT_BOOST_MIN : 0;
 231         sg_cpu->iowait_boost_pending = set_iowait_boost;
 232
 233         return true;
 234 }
 235
 236 /**
 237  * sugov_iowait_boost() - Updates the IO boost status of a CPU.
 238  * @sg_cpu: the sugov data for the CPU to boost
 239  * @time: the update time from the caller
 240  * @flags: SCHED_CPUFREQ_IOWAIT if the task is waking up after an IO wait
 241  *
 242  * Each time a task wakes up after an IO operation, the CPU utilization can be
 243  * boosted to a certain utilization which doubles at each "frequent and
 244  * successive" wakeup from IO, ranging from IOWAIT_BOOST_MIN to the utilization
 245  * of the maximum OPP.
 246  *
 247  * To keep doubling, an IO boost has to be requested at least once per tick,
 248  * otherwise we restart from the utilization of the minimum OPP.
 249  */
 250 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
 251                                unsigned int flags)
 252 {
 253         bool set_iowait_boost = flags & SCHED_CPUFREQ_IOWAIT;
 254
 255         /* Reset boost if the CPU appears to have been idle enough */
 256         if (sg_cpu->iowait_boost &&
 257             sugov_iowait_reset(sg_cpu, time, set_iowait_boost))
 258                 return;
 259
 260         /* Boost only tasks waking up after IO */
 261         if (!set_iowait_boost)
 262                 return;
 263
 264         /* Ensure boost doubles only one time at each request */
 265         if (sg_cpu->iowait_boost_pending)
 266                 return;
 267         sg_cpu->iowait_boost_pending = true;
 268
 269         /* Double the boost at each request */
 270         if (sg_cpu->iowait_boost) {
 271                 sg_cpu->iowait_boost =
 272                         min_t(unsigned int, sg_cpu->iowait_boost << 1, SCHED_CAPACITY_SCALE);
 273                 return;
 274         }
 275
 276         /* First wakeup after IO: start with minimum boost */
 277         sg_cpu->iowait_boost = IOWAIT_BOOST_MIN;
 278 }
 279
 280 /**
 281  * sugov_iowait_apply() - Apply the IO boost to a CPU.
 282  * @sg_cpu: the sugov data for the cpu to boost
 283  * @time: the update time from the caller
 284  * @max_cap: the max CPU capacity
 285  *
 286  * A CPU running a task which woken up after an IO operation can have its
 287  * utilization boosted to speed up the completion of those IO operations.
 288  * The IO boost value is increased each time a task wakes up from IO, in
 289  * sugov_iowait_apply(), and it's instead decreased by this function,
 290  * each time an increase has not been requested (!iowait_boost_pending).
 291  *
 292  * A CPU which also appears to have been idle for at least one tick has also
 293  * its IO boost utilization reset.
 294  *
 295  * This mechanism is designed to boost high frequently IO waiting tasks, while
 296  * being more conservative on tasks which does sporadic IO operations.
 297  */
 298 static unsigned long sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
 299                                unsigned long max_cap)
 300 {
 301         /* No boost currently required */
 302         if (!sg_cpu->iowait_boost)
 303                 return 0;
 304
 305         /* Reset boost if the CPU appears to have been idle enough */
 306         if (sugov_iowait_reset(sg_cpu, time, false))
 307                 return 0;
 308
 309         if (!sg_cpu->iowait_boost_pending) {
 310                 /*
 311                  * No boost pending; reduce the boost value.
 312                  */
 313                 sg_cpu->iowait_boost >>= 1;
 314                 if (sg_cpu->iowait_boost < IOWAIT_BOOST_MIN) {
 315                         sg_cpu->iowait_boost = 0;
 316                         return 0;
 317                 }
 318         }
 319
 320         sg_cpu->iowait_boost_pending = false;
 321
 322         /*
 323          * sg_cpu->util is already in capacity scale; convert iowait_boost
 324          * into the same scale so we can compare.
 325          */
 326         return (sg_cpu->iowait_boost * max_cap) >> SCHED_CAPACITY_SHIFT;
 327 }
 328
 329 #ifdef CONFIG_NO_HZ_COMMON
 330 static bool sugov_hold_freq(struct sugov_cpu *sg_cpu)
 331 {
 332         unsigned long idle_calls;
 333         bool ret;
 334
 335         /*
 336          * The heuristics in this function is for the fair class. For SCX, the
 337          * performance target comes directly from the BPF scheduler. Let's just
 338          * follow it.
 339          */
 340         if (scx_switched_all())
 341                 return false;
 342
 343         /* if capped by uclamp_max, always update to be in compliance */
 344         if (uclamp_rq_is_capped(cpu_rq(sg_cpu->cpu)))
 345                 return false;
 346
 347         /*
 348          * Maintain the frequency if the CPU has not been idle recently, as
 349          * reduction is likely to be premature.
 350          */
 351         idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
 352         ret = idle_calls == sg_cpu->saved_idle_calls;
 353
 354         sg_cpu->saved_idle_calls = idle_calls;
 355         return ret;
 356 }
 357 #else
 358 static inline bool sugov_hold_freq(struct sugov_cpu *sg_cpu) { return false; }
 359 #endif /* CONFIG_NO_HZ_COMMON */
 360
 361 /*
 362  * Make sugov_should_update_freq() ignore the rate limit when DL
 363  * has increased the utilization.
 364  */
 365 static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu)
 366 {
 367         if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_min)
 368                 sg_cpu->sg_policy->limits_changed = true;
 369 }
 370
 371 static inline bool sugov_update_single_common(struct sugov_cpu *sg_cpu,
 372                                               u64 time, unsigned long max_cap,
 373                                               unsigned int flags)
 374 {
 375         unsigned long boost;
 376
 377         sugov_iowait_boost(sg_cpu, time, flags);
 378         sg_cpu->last_update = time;
 379
 380         ignore_dl_rate_limit(sg_cpu);
 381
 382         if (!sugov_should_update_freq(sg_cpu->sg_policy, time))
 383                 return false;
 384
 385         boost = sugov_iowait_apply(sg_cpu, time, max_cap);
 386         sugov_get_util(sg_cpu, boost);
 387
 388         return true;
 389 }
 390
 391 static void sugov_update_single_freq(struct update_util_data *hook, u64 time,
 392                                      unsigned int flags)
 393 {
 394         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 395         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 396         unsigned int cached_freq = sg_policy->cached_raw_freq;
 397         unsigned long max_cap;
 398         unsigned int next_f;
 399
 400         max_cap = arch_scale_cpu_capacity(sg_cpu->cpu);
 401
 402         if (!sugov_update_single_common(sg_cpu, time, max_cap, flags))
 403                 return;
 404
 405         next_f = get_next_freq(sg_policy, sg_cpu->util, max_cap);
 406
 407         if (sugov_hold_freq(sg_cpu) && next_f < sg_policy->next_freq &&
 408             !sg_policy->need_freq_update) {
 409                 next_f = sg_policy->next_freq;
 410
 411                 /* Restore cached freq as next_freq has changed */
 412                 sg_policy->cached_raw_freq = cached_freq;
 413         }
 414
 415         if (!sugov_update_next_freq(sg_policy, time, next_f))
 416                 return;
 417
 418         /*
 419          * This code runs under rq->lock for the target CPU, so it won't run
 420          * concurrently on two different CPUs for the same target and it is not
 421          * necessary to acquire the lock in the fast switch case.
 422          */
 423         if (sg_policy->policy->fast_switch_enabled) {
 424                 cpufreq_driver_fast_switch(sg_policy->policy, next_f);
 425         } else {
 426                 raw_spin_lock(&sg_policy->update_lock);
 427                 sugov_deferred_update(sg_policy);
 428                 raw_spin_unlock(&sg_policy->update_lock);
 429         }
 430 }
 431
 432 static void sugov_update_single_perf(struct update_util_data *hook, u64 time,
 433                                      unsigned int flags)
 434 {
 435         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 436         unsigned long prev_util = sg_cpu->util;
 437         unsigned long max_cap;
 438
 439         /*
 440          * Fall back to the "frequency" path if frequency invariance is not
 441          * supported, because the direct mapping between the utilization and
 442          * the performance levels depends on the frequency invariance.
 443          */
 444         if (!arch_scale_freq_invariant()) {
 445                 sugov_update_single_freq(hook, time, flags);
 446                 return;
 447         }
 448
 449         max_cap = arch_scale_cpu_capacity(sg_cpu->cpu);
 450
 451         if (!sugov_update_single_common(sg_cpu, time, max_cap, flags))
 452                 return;
 453
 454         if (sugov_hold_freq(sg_cpu) && sg_cpu->util < prev_util)
 455                 sg_cpu->util = prev_util;
 456
 457         cpufreq_driver_adjust_perf(sg_cpu->cpu, sg_cpu->bw_min,
 458                                    sg_cpu->util, max_cap);
 459
 460         sg_cpu->sg_policy->last_freq_update_time = time;
 461 }
 462
 463 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
 464 {
 465         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 466         struct cpufreq_policy *policy = sg_policy->policy;
 467         unsigned long util = 0, max_cap;
 468         unsigned int j;
 469
 470         max_cap = arch_scale_cpu_capacity(sg_cpu->cpu);
 471
 472         for_each_cpu(j, policy->cpus) {
 473                 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
 474                 unsigned long boost;
 475
 476                 boost = sugov_iowait_apply(j_sg_cpu, time, max_cap);
 477                 sugov_get_util(j_sg_cpu, boost);
 478
 479                 util = max(j_sg_cpu->util, util);
 480         }
 481
 482         return get_next_freq(sg_policy, util, max_cap);
 483 }
 484
 485 static void
 486 sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
 487 {
 488         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 489         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 490         unsigned int next_f;
 491
 492         raw_spin_lock(&sg_policy->update_lock);
 493
 494         sugov_iowait_boost(sg_cpu, time, flags);
 495         sg_cpu->last_update = time;
 496
 497         ignore_dl_rate_limit(sg_cpu);
 498
 499         if (sugov_should_update_freq(sg_policy, time)) {
 500                 next_f = sugov_next_freq_shared(sg_cpu, time);
 501
 502                 if (!sugov_update_next_freq(sg_policy, time, next_f))
 503                         goto unlock;
 504
 505                 if (sg_policy->policy->fast_switch_enabled)
 506                         cpufreq_driver_fast_switch(sg_policy->policy, next_f);
 507                 else
 508                         sugov_deferred_update(sg_policy);
 509         }
 510 unlock:
 511         raw_spin_unlock(&sg_policy->update_lock);
 512 }
 513
 514 static void sugov_work(struct kthread_work *work)
 515 {
 516         struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
 517         unsigned int freq;
 518         unsigned long flags;
 519
 520         /*
 521          * Hold sg_policy->update_lock shortly to handle the case where:
 522          * in case sg_policy->next_freq is read here, and then updated by
 523          * sugov_deferred_update() just before work_in_progress is set to false
 524          * here, we may miss queueing the new update.
 525          *
 526          * Note: If a work was queued after the update_lock is released,
 527          * sugov_work() will just be called again by kthread_work code; and the
 528          * request will be proceed before the sugov thread sleeps.
 529          */
 530         raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
 531         freq = sg_policy->next_freq;
 532         sg_policy->work_in_progress = false;
 533         raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
 534
 535         mutex_lock(&sg_policy->work_lock);
 536         __cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
 537         mutex_unlock(&sg_policy->work_lock);
 538 }
 539
 540 static void sugov_irq_work(struct irq_work *irq_work)
 541 {
 542         struct sugov_policy *sg_policy;
 543
 544         sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
 545
 546         kthread_queue_work(&sg_policy->worker, &sg_policy->work);
 547 }
 548
 549 /************************** sysfs interface ************************/
 550
 551 static struct sugov_tunables *global_tunables;
 552 static DEFINE_MUTEX(global_tunables_lock);
 553
 554 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
 555 {
 556         return container_of(attr_set, struct sugov_tunables, attr_set);
 557 }
 558
 559 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
 560 {
 561         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 562
 563         return sprintf(buf, "%u\n", tunables->rate_limit_us);
 564 }
 565
 566 static ssize_t
 567 rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
 568 {
 569         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 570         struct sugov_policy *sg_policy;
 571         unsigned int rate_limit_us;
 572
 573         if (kstrtouint(buf, 10, &rate_limit_us))
 574                 return -EINVAL;
 575
 576         tunables->rate_limit_us = rate_limit_us;
 577
 578         list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
 579                 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
 580
 581         return count;
 582 }
 583
 584 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
 585
 586 static struct attribute *sugov_attrs[] = {
 587         &rate_limit_us.attr,
 588         NULL
 589 };
 590 ATTRIBUTE_GROUPS(sugov);
 591
 592 static void sugov_tunables_free(struct kobject *kobj)
 593 {
 594         struct gov_attr_set *attr_set = to_gov_attr_set(kobj);
 595
 596         kfree(to_sugov_tunables(attr_set));
 597 }
 598
 599 static const struct kobj_type sugov_tunables_ktype = {
 600         .default_groups = sugov_groups,
 601         .sysfs_ops = &governor_sysfs_ops,
 602         .release = &sugov_tunables_free,
 603 };
 604
 605 /********************** cpufreq governor interface *********************/
 606
 607 #ifdef CONFIG_ENERGY_MODEL
 608 static void rebuild_sd_workfn(struct work_struct *work)
 609 {
 610         rebuild_sched_domains_energy();
 611 }
 612
 613 static DECLARE_WORK(rebuild_sd_work, rebuild_sd_workfn);
 614
 615 /*
 616  * EAS shouldn't be attempted without sugov, so rebuild the sched_domains
 617  * on governor changes to make sure the scheduler knows about it.
 618  */
 619 static void sugov_eas_rebuild_sd(void)
 620 {
 621         /*
 622          * When called from the cpufreq_register_driver() path, the
 623          * cpu_hotplug_lock is already held, so use a work item to
 624          * avoid nested locking in rebuild_sched_domains().
 625          */
 626         schedule_work(&rebuild_sd_work);
 627 }
 628 #else
 629 static inline void sugov_eas_rebuild_sd(void) { };
 630 #endif
 631
 632 struct cpufreq_governor schedutil_gov;
 633
 634 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
 635 {
 636         struct sugov_policy *sg_policy;
 637
 638         sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
 639         if (!sg_policy)
 640                 return NULL;
 641
 642         sg_policy->policy = policy;
 643         raw_spin_lock_init(&sg_policy->update_lock);
 644         return sg_policy;
 645 }
 646
 647 static void sugov_policy_free(struct sugov_policy *sg_policy)
 648 {
 649         kfree(sg_policy);
 650 }
 651
 652 static int sugov_kthread_create(struct sugov_policy *sg_policy)
 653 {
 654         struct task_struct *thread;
 655         struct sched_attr attr = {
 656                 .size           = sizeof(struct sched_attr),
 657                 .sched_policy   = SCHED_DEADLINE,
 658                 .sched_flags    = SCHED_FLAG_SUGOV,
 659                 .sched_nice     = 0,
 660                 .sched_priority = 0,
 661                 /*
 662                  * Fake (unused) bandwidth; workaround to "fix"
 663                  * priority inheritance.
 664                  */
 665                 .sched_runtime  = NSEC_PER_MSEC,
 666                 .sched_deadline = 10 * NSEC_PER_MSEC,
 667                 .sched_period   = 10 * NSEC_PER_MSEC,
 668         };
 669         struct cpufreq_policy *policy = sg_policy->policy;
 670         int ret;
 671
 672         /* kthread only required for slow path */
 673         if (policy->fast_switch_enabled)
 674                 return 0;
 675
 676         kthread_init_work(&sg_policy->work, sugov_work);
 677         kthread_init_worker(&sg_policy->worker);
 678         thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
 679                                 "sugov:%d",
 680                                 cpumask_first(policy->related_cpus));
 681         if (IS_ERR(thread)) {
 682                 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
 683                 return PTR_ERR(thread);
 684         }
 685
 686         ret = sched_setattr_nocheck(thread, &attr);
 687         if (ret) {
 688                 kthread_stop(thread);
 689                 pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
 690                 return ret;
 691         }
 692
 693         sg_policy->thread = thread;
 694         kthread_bind_mask(thread, policy->related_cpus);
 695         init_irq_work(&sg_policy->irq_work, sugov_irq_work);
 696         mutex_init(&sg_policy->work_lock);
 697
 698         wake_up_process(thread);
 699
 700         return 0;
 701 }
 702
 703 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
 704 {
 705         /* kthread only required for slow path */
 706         if (sg_policy->policy->fast_switch_enabled)
 707                 return;
 708
 709         kthread_flush_worker(&sg_policy->worker);
 710         kthread_stop(sg_policy->thread);
 711         mutex_destroy(&sg_policy->work_lock);
 712 }
 713
 714 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
 715 {
 716         struct sugov_tunables *tunables;
 717
 718         tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
 719         if (tunables) {
 720                 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
 721                 if (!have_governor_per_policy())
 722                         global_tunables = tunables;
 723         }
 724         return tunables;
 725 }
 726
 727 static void sugov_clear_global_tunables(void)
 728 {
 729         if (!have_governor_per_policy())
 730                 global_tunables = NULL;
 731 }
 732
 733 static int sugov_init(struct cpufreq_policy *policy)
 734 {
 735         struct sugov_policy *sg_policy;
 736         struct sugov_tunables *tunables;
 737         int ret = 0;
 738
 739         /* State should be equivalent to EXIT */
 740         if (policy->governor_data)
 741                 return -EBUSY;
 742
 743         cpufreq_enable_fast_switch(policy);
 744
 745         sg_policy = sugov_policy_alloc(policy);
 746         if (!sg_policy) {
 747                 ret = -ENOMEM;
 748                 goto disable_fast_switch;
 749         }
 750
 751         ret = sugov_kthread_create(sg_policy);
 752         if (ret)
 753                 goto free_sg_policy;
 754
 755         mutex_lock(&global_tunables_lock);
 756
 757         if (global_tunables) {
 758                 if (WARN_ON(have_governor_per_policy())) {
 759                         ret = -EINVAL;
 760                         goto stop_kthread;
 761                 }
 762                 policy->governor_data = sg_policy;
 763                 sg_policy->tunables = global_tunables;
 764
 765                 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
 766                 goto out;
 767         }
 768
 769         tunables = sugov_tunables_alloc(sg_policy);
 770         if (!tunables) {
 771                 ret = -ENOMEM;
 772                 goto stop_kthread;
 773         }
 774
 775         tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
 776
 777         policy->governor_data = sg_policy;
 778         sg_policy->tunables = tunables;
 779
 780         ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
 781                                    get_governor_parent_kobj(policy), "%s",
 782                                    schedutil_gov.name);
 783         if (ret)
 784                 goto fail;
 785
 786 out:
 787         sugov_eas_rebuild_sd();
 788         mutex_unlock(&global_tunables_lock);
 789         return 0;
 790
 791 fail:
 792         kobject_put(&tunables->attr_set.kobj);
 793         policy->governor_data = NULL;
 794         sugov_clear_global_tunables();
 795
 796 stop_kthread:
 797         sugov_kthread_stop(sg_policy);
 798         mutex_unlock(&global_tunables_lock);
 799
 800 free_sg_policy:
 801         sugov_policy_free(sg_policy);
 802
 803 disable_fast_switch:
 804         cpufreq_disable_fast_switch(policy);
 805
 806         pr_err("initialization failed (error %d)\n", ret);
 807         return ret;
 808 }
 809
 810 static void sugov_exit(struct cpufreq_policy *policy)
 811 {
 812         struct sugov_policy *sg_policy = policy->governor_data;
 813         struct sugov_tunables *tunables = sg_policy->tunables;
 814         unsigned int count;
 815
 816         mutex_lock(&global_tunables_lock);
 817
 818         count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
 819         policy->governor_data = NULL;
 820         if (!count)
 821                 sugov_clear_global_tunables();
 822
 823         mutex_unlock(&global_tunables_lock);
 824
 825         sugov_kthread_stop(sg_policy);
 826         sugov_policy_free(sg_policy);
 827         cpufreq_disable_fast_switch(policy);
 828
 829         sugov_eas_rebuild_sd();
 830 }
 831
 832 static int sugov_start(struct cpufreq_policy *policy)
 833 {
 834         struct sugov_policy *sg_policy = policy->governor_data;
 835         void (*uu)(struct update_util_data *data, u64 time, unsigned int flags);
 836         unsigned int cpu;
 837
 838         sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
 839         sg_policy->last_freq_update_time        = 0;
 840         sg_policy->next_freq                    = 0;
 841         sg_policy->work_in_progress             = false;
 842         sg_policy->limits_changed               = false;
 843         sg_policy->cached_raw_freq              = 0;
 844
 845         sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
 846
 847         if (policy_is_shared(policy))
 848                 uu = sugov_update_shared;
 849         else if (policy->fast_switch_enabled && cpufreq_driver_has_adjust_perf())
 850                 uu = sugov_update_single_perf;
 851         else
 852                 uu = sugov_update_single_freq;
 853
 854         for_each_cpu(cpu, policy->cpus) {
 855                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 856
 857                 memset(sg_cpu, 0, sizeof(*sg_cpu));
 858                 sg_cpu->cpu = cpu;
 859                 sg_cpu->sg_policy = sg_policy;
 860                 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util, uu);
 861         }
 862         return 0;
 863 }
 864
 865 static void sugov_stop(struct cpufreq_policy *policy)
 866 {
 867         struct sugov_policy *sg_policy = policy->governor_data;
 868         unsigned int cpu;
 869
 870         for_each_cpu(cpu, policy->cpus)
 871                 cpufreq_remove_update_util_hook(cpu);
 872
 873         synchronize_rcu();
 874
 875         if (!policy->fast_switch_enabled) {
 876                 irq_work_sync(&sg_policy->irq_work);
 877                 kthread_cancel_work_sync(&sg_policy->work);
 878         }
 879 }
 880
 881 static void sugov_limits(struct cpufreq_policy *policy)
 882 {
 883         struct sugov_policy *sg_policy = policy->governor_data;
 884
 885         if (!policy->fast_switch_enabled) {
 886                 mutex_lock(&sg_policy->work_lock);
 887                 cpufreq_policy_apply_limits(policy);
 888                 mutex_unlock(&sg_policy->work_lock);
 889         }
 890
 891         sg_policy->limits_changed = true;
 892 }
 893
 894 struct cpufreq_governor schedutil_gov = {
 895         .name                   = "schedutil",
 896         .owner                  = THIS_MODULE,
 897         .flags                  = CPUFREQ_GOV_DYNAMIC_SWITCHING,
 898         .init                   = sugov_init,
 899         .exit                   = sugov_exit,
 900         .start                  = sugov_start,
 901         .stop                   = sugov_stop,
 902         .limits                 = sugov_limits,
 903 };
 904
 905 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
 906 struct cpufreq_governor *cpufreq_default_governor(void)
 907 {
 908         return &schedutil_gov;
 909 }
 910 #endif
 911
 912 cpufreq_governor_init(schedutil_gov);