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;    /* For shared policies */
  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           bw_dl;
  57         unsigned long           max;
  58
  59         /* The field below is for single-CPU policies only: */
  60 #ifdef CONFIG_NO_HZ_COMMON
  61         unsigned long           saved_idle_calls;
  62 #endif
  63 };
  64
  65 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
  66
  67 /************************ Governor internals ***********************/
  68
  69 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
  70 {
  71         s64 delta_ns;
  72
  73         /*
  74          * Since cpufreq_update_util() is called with rq->lock held for
  75          * the @target_cpu, our per-CPU data is fully serialized.
  76          *
  77          * However, drivers cannot in general deal with cross-CPU
  78          * requests, so while get_next_freq() will work, our
  79          * sugov_update_commit() call may not for the fast switching platforms.
  80          *
  81          * Hence stop here for remote requests if they aren't supported
  82          * by the hardware, as calculating the frequency is pointless if
  83          * we cannot in fact act on it.
  84          *
  85          * This is needed on the slow switching platforms too to prevent CPUs
  86          * going offline from leaving stale IRQ work items behind.
  87          */
  88         if (!cpufreq_this_cpu_can_update(sg_policy->policy))
  89                 return false;
  90
  91         if (unlikely(sg_policy->limits_changed)) {
  92                 sg_policy->limits_changed = false;
  93                 sg_policy->need_freq_update = true;
  94                 return true;
  95         }
  96
  97         delta_ns = time - sg_policy->last_freq_update_time;
  98
  99         return delta_ns >= sg_policy->freq_update_delay_ns;
 100 }
 101
 102 static bool sugov_update_next_freq(struct sugov_policy *sg_policy, u64 time,
 103                                    unsigned int next_freq)
 104 {
 105         if (sg_policy->next_freq == next_freq &&
 106             !cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS))
 107                 return false;
 108
 109         sg_policy->next_freq = next_freq;
 110         sg_policy->last_freq_update_time = time;
 111
 112         return true;
 113 }
 114
 115 static void sugov_fast_switch(struct sugov_policy *sg_policy, u64 time,
 116                               unsigned int next_freq)
 117 {
 118         if (sugov_update_next_freq(sg_policy, time, next_freq))
 119                 cpufreq_driver_fast_switch(sg_policy->policy, next_freq);
 120 }
 121
 122 static void sugov_deferred_update(struct sugov_policy *sg_policy, u64 time,
 123                                   unsigned int next_freq)
 124 {
 125         if (!sugov_update_next_freq(sg_policy, time, next_freq))
 126                 return;
 127
 128         if (!sg_policy->work_in_progress) {
 129                 sg_policy->work_in_progress = true;
 130                 irq_work_queue(&sg_policy->irq_work);
 131         }
 132 }
 133
 134 /**
 135  * get_next_freq - Compute a new frequency for a given cpufreq policy.
 136  * @sg_policy: schedutil policy object to compute the new frequency for.
 137  * @util: Current CPU utilization.
 138  * @max: CPU capacity.
 139  *
 140  * If the utilization is frequency-invariant, choose the new frequency to be
 141  * proportional to it, that is
 142  *
 143  * next_freq = C * max_freq * util / max
 144  *
 145  * Otherwise, approximate the would-be frequency-invariant utilization by
 146  * util_raw * (curr_freq / max_freq) which leads to
 147  *
 148  * next_freq = C * curr_freq * util_raw / max
 149  *
 150  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
 151  *
 152  * The lowest driver-supported frequency which is equal or greater than the raw
 153  * next_freq (as calculated above) is returned, subject to policy min/max and
 154  * cpufreq driver limitations.
 155  */
 156 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
 157                                   unsigned long util, unsigned long max)
 158 {
 159         struct cpufreq_policy *policy = sg_policy->policy;
 160         unsigned int freq = arch_scale_freq_invariant() ?
 161                                 policy->cpuinfo.max_freq : policy->cur;
 162
 163         freq = map_util_freq(util, freq, max);
 164
 165         if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update &&
 166             !cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS))
 167                 return sg_policy->next_freq;
 168
 169         sg_policy->need_freq_update = false;
 170         sg_policy->cached_raw_freq = freq;
 171         return cpufreq_driver_resolve_freq(policy, freq);
 172 }
 173
 174 /*
 175  * This function computes an effective utilization for the given CPU, to be
 176  * used for frequency selection given the linear relation: f = u * f_max.
 177  *
 178  * The scheduler tracks the following metrics:
 179  *
 180  *   cpu_util_{cfs,rt,dl,irq}()
 181  *   cpu_bw_dl()
 182  *
 183  * Where the cfs,rt and dl util numbers are tracked with the same metric and
 184  * synchronized windows and are thus directly comparable.
 185  *
 186  * The cfs,rt,dl utilization are the running times measured with rq->clock_task
 187  * which excludes things like IRQ and steal-time. These latter are then accrued
 188  * in the irq utilization.
 189  *
 190  * The DL bandwidth number otoh is not a measured metric but a value computed
 191  * based on the task model parameters and gives the minimal utilization
 192  * required to meet deadlines.
 193  */
 194 unsigned long schedutil_cpu_util(int cpu, unsigned long util_cfs,
 195                                  unsigned long max, enum schedutil_type type,
 196                                  struct task_struct *p)
 197 {
 198         unsigned long dl_util, util, irq;
 199         struct rq *rq = cpu_rq(cpu);
 200
 201         if (!uclamp_is_used() &&
 202             type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
 203                 return max;
 204         }
 205
 206         /*
 207          * Early check to see if IRQ/steal time saturates the CPU, can be
 208          * because of inaccuracies in how we track these -- see
 209          * update_irq_load_avg().
 210          */
 211         irq = cpu_util_irq(rq);
 212         if (unlikely(irq >= max))
 213                 return max;
 214
 215         /*
 216          * Because the time spend on RT/DL tasks is visible as 'lost' time to
 217          * CFS tasks and we use the same metric to track the effective
 218          * utilization (PELT windows are synchronized) we can directly add them
 219          * to obtain the CPU's actual utilization.
 220          *
 221          * CFS and RT utilization can be boosted or capped, depending on
 222          * utilization clamp constraints requested by currently RUNNABLE
 223          * tasks.
 224          * When there are no CFS RUNNABLE tasks, clamps are released and
 225          * frequency will be gracefully reduced with the utilization decay.
 226          */
 227         util = util_cfs + cpu_util_rt(rq);
 228         if (type == FREQUENCY_UTIL)
 229                 util = uclamp_rq_util_with(rq, util, p);
 230
 231         dl_util = cpu_util_dl(rq);
 232
 233         /*
 234          * For frequency selection we do not make cpu_util_dl() a permanent part
 235          * of this sum because we want to use cpu_bw_dl() later on, but we need
 236          * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
 237          * that we select f_max when there is no idle time.
 238          *
 239          * NOTE: numerical errors or stop class might cause us to not quite hit
 240          * saturation when we should -- something for later.
 241          */
 242         if (util + dl_util >= max)
 243                 return max;
 244
 245         /*
 246          * OTOH, for energy computation we need the estimated running time, so
 247          * include util_dl and ignore dl_bw.
 248          */
 249         if (type == ENERGY_UTIL)
 250                 util += dl_util;
 251
 252         /*
 253          * There is still idle time; further improve the number by using the
 254          * irq metric. Because IRQ/steal time is hidden from the task clock we
 255          * need to scale the task numbers:
 256          *
 257          *              max - irq
 258          *   U' = irq + --------- * U
 259          *                 max
 260          */
 261         util = scale_irq_capacity(util, irq, max);
 262         util += irq;
 263
 264         /*
 265          * Bandwidth required by DEADLINE must always be granted while, for
 266          * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
 267          * to gracefully reduce the frequency when no tasks show up for longer
 268          * periods of time.
 269          *
 270          * Ideally we would like to set bw_dl as min/guaranteed freq and util +
 271          * bw_dl as requested freq. However, cpufreq is not yet ready for such
 272          * an interface. So, we only do the latter for now.
 273          */
 274         if (type == FREQUENCY_UTIL)
 275                 util += cpu_bw_dl(rq);
 276
 277         return min(max, util);
 278 }
 279
 280 static unsigned long sugov_get_util(struct sugov_cpu *sg_cpu)
 281 {
 282         struct rq *rq = cpu_rq(sg_cpu->cpu);
 283         unsigned long util = cpu_util_cfs(rq);
 284         unsigned long max = arch_scale_cpu_capacity(sg_cpu->cpu);
 285
 286         sg_cpu->max = max;
 287         sg_cpu->bw_dl = cpu_bw_dl(rq);
 288
 289         return schedutil_cpu_util(sg_cpu->cpu, util, max, FREQUENCY_UTIL, NULL);
 290 }
 291
 292 /**
 293  * sugov_iowait_reset() - Reset the IO boost status of a CPU.
 294  * @sg_cpu: the sugov data for the CPU to boost
 295  * @time: the update time from the caller
 296  * @set_iowait_boost: true if an IO boost has been requested
 297  *
 298  * The IO wait boost of a task is disabled after a tick since the last update
 299  * of a CPU. If a new IO wait boost is requested after more then a tick, then
 300  * we enable the boost starting from IOWAIT_BOOST_MIN, which improves energy
 301  * efficiency by ignoring sporadic wakeups from IO.
 302  */
 303 static bool sugov_iowait_reset(struct sugov_cpu *sg_cpu, u64 time,
 304                                bool set_iowait_boost)
 305 {
 306         s64 delta_ns = time - sg_cpu->last_update;
 307
 308         /* Reset boost only if a tick has elapsed since last request */
 309         if (delta_ns <= TICK_NSEC)
 310                 return false;
 311
 312         sg_cpu->iowait_boost = set_iowait_boost ? IOWAIT_BOOST_MIN : 0;
 313         sg_cpu->iowait_boost_pending = set_iowait_boost;
 314
 315         return true;
 316 }
 317
 318 /**
 319  * sugov_iowait_boost() - Updates the IO boost status of a CPU.
 320  * @sg_cpu: the sugov data for the CPU to boost
 321  * @time: the update time from the caller
 322  * @flags: SCHED_CPUFREQ_IOWAIT if the task is waking up after an IO wait
 323  *
 324  * Each time a task wakes up after an IO operation, the CPU utilization can be
 325  * boosted to a certain utilization which doubles at each "frequent and
 326  * successive" wakeup from IO, ranging from IOWAIT_BOOST_MIN to the utilization
 327  * of the maximum OPP.
 328  *
 329  * To keep doubling, an IO boost has to be requested at least once per tick,
 330  * otherwise we restart from the utilization of the minimum OPP.
 331  */
 332 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
 333                                unsigned int flags)
 334 {
 335         bool set_iowait_boost = flags & SCHED_CPUFREQ_IOWAIT;
 336
 337         /* Reset boost if the CPU appears to have been idle enough */
 338         if (sg_cpu->iowait_boost &&
 339             sugov_iowait_reset(sg_cpu, time, set_iowait_boost))
 340                 return;
 341
 342         /* Boost only tasks waking up after IO */
 343         if (!set_iowait_boost)
 344                 return;
 345
 346         /* Ensure boost doubles only one time at each request */
 347         if (sg_cpu->iowait_boost_pending)
 348                 return;
 349         sg_cpu->iowait_boost_pending = true;
 350
 351         /* Double the boost at each request */
 352         if (sg_cpu->iowait_boost) {
 353                 sg_cpu->iowait_boost =
 354                         min_t(unsigned int, sg_cpu->iowait_boost << 1, SCHED_CAPACITY_SCALE);
 355                 return;
 356         }
 357
 358         /* First wakeup after IO: start with minimum boost */
 359         sg_cpu->iowait_boost = IOWAIT_BOOST_MIN;
 360 }
 361
 362 /**
 363  * sugov_iowait_apply() - Apply the IO boost to a CPU.
 364  * @sg_cpu: the sugov data for the cpu to boost
 365  * @time: the update time from the caller
 366  * @util: the utilization to (eventually) boost
 367  * @max: the maximum value the utilization can be boosted to
 368  *
 369  * A CPU running a task which woken up after an IO operation can have its
 370  * utilization boosted to speed up the completion of those IO operations.
 371  * The IO boost value is increased each time a task wakes up from IO, in
 372  * sugov_iowait_apply(), and it's instead decreased by this function,
 373  * each time an increase has not been requested (!iowait_boost_pending).
 374  *
 375  * A CPU which also appears to have been idle for at least one tick has also
 376  * its IO boost utilization reset.
 377  *
 378  * This mechanism is designed to boost high frequently IO waiting tasks, while
 379  * being more conservative on tasks which does sporadic IO operations.
 380  */
 381 static unsigned long sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
 382                                         unsigned long util, unsigned long max)
 383 {
 384         unsigned long boost;
 385
 386         /* No boost currently required */
 387         if (!sg_cpu->iowait_boost)
 388                 return util;
 389
 390         /* Reset boost if the CPU appears to have been idle enough */
 391         if (sugov_iowait_reset(sg_cpu, time, false))
 392                 return util;
 393
 394         if (!sg_cpu->iowait_boost_pending) {
 395                 /*
 396                  * No boost pending; reduce the boost value.
 397                  */
 398                 sg_cpu->iowait_boost >>= 1;
 399                 if (sg_cpu->iowait_boost < IOWAIT_BOOST_MIN) {
 400                         sg_cpu->iowait_boost = 0;
 401                         return util;
 402                 }
 403         }
 404
 405         sg_cpu->iowait_boost_pending = false;
 406
 407         /*
 408          * @util is already in capacity scale; convert iowait_boost
 409          * into the same scale so we can compare.
 410          */
 411         boost = (sg_cpu->iowait_boost * max) >> SCHED_CAPACITY_SHIFT;
 412         return max(boost, util);
 413 }
 414
 415 #ifdef CONFIG_NO_HZ_COMMON
 416 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
 417 {
 418         unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
 419         bool ret = idle_calls == sg_cpu->saved_idle_calls;
 420
 421         sg_cpu->saved_idle_calls = idle_calls;
 422         return ret;
 423 }
 424 #else
 425 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
 426 #endif /* CONFIG_NO_HZ_COMMON */
 427
 428 /*
 429  * Make sugov_should_update_freq() ignore the rate limit when DL
 430  * has increased the utilization.
 431  */
 432 static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu, struct sugov_policy *sg_policy)
 433 {
 434         if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_dl)
 435                 sg_policy->limits_changed = true;
 436 }
 437
 438 static void sugov_update_single(struct update_util_data *hook, u64 time,
 439                                 unsigned int flags)
 440 {
 441         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 442         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 443         unsigned long util, max;
 444         unsigned int next_f;
 445         bool busy;
 446         unsigned int cached_freq = sg_policy->cached_raw_freq;
 447
 448         sugov_iowait_boost(sg_cpu, time, flags);
 449         sg_cpu->last_update = time;
 450
 451         ignore_dl_rate_limit(sg_cpu, sg_policy);
 452
 453         if (!sugov_should_update_freq(sg_policy, time))
 454                 return;
 455
 456         /* Limits may have changed, don't skip frequency update */
 457         busy = !sg_policy->need_freq_update && sugov_cpu_is_busy(sg_cpu);
 458
 459         util = sugov_get_util(sg_cpu);
 460         max = sg_cpu->max;
 461         util = sugov_iowait_apply(sg_cpu, time, util, max);
 462         next_f = get_next_freq(sg_policy, util, max);
 463         /*
 464          * Do not reduce the frequency if the CPU has not been idle
 465          * recently, as the reduction is likely to be premature then.
 466          */
 467         if (busy && next_f < sg_policy->next_freq) {
 468                 next_f = sg_policy->next_freq;
 469
 470                 /* Restore cached freq as next_freq has changed */
 471                 sg_policy->cached_raw_freq = cached_freq;
 472         }
 473
 474         /*
 475          * This code runs under rq->lock for the target CPU, so it won't run
 476          * concurrently on two different CPUs for the same target and it is not
 477          * necessary to acquire the lock in the fast switch case.
 478          */
 479         if (sg_policy->policy->fast_switch_enabled) {
 480                 sugov_fast_switch(sg_policy, time, next_f);
 481         } else {
 482                 raw_spin_lock(&sg_policy->update_lock);
 483                 sugov_deferred_update(sg_policy, time, next_f);
 484                 raw_spin_unlock(&sg_policy->update_lock);
 485         }
 486 }
 487
 488 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
 489 {
 490         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 491         struct cpufreq_policy *policy = sg_policy->policy;
 492         unsigned long util = 0, max = 1;
 493         unsigned int j;
 494
 495         for_each_cpu(j, policy->cpus) {
 496                 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
 497                 unsigned long j_util, j_max;
 498
 499                 j_util = sugov_get_util(j_sg_cpu);
 500                 j_max = j_sg_cpu->max;
 501                 j_util = sugov_iowait_apply(j_sg_cpu, time, j_util, j_max);
 502
 503                 if (j_util * max > j_max * util) {
 504                         util = j_util;
 505                         max = j_max;
 506                 }
 507         }
 508
 509         return get_next_freq(sg_policy, util, max);
 510 }
 511
 512 static void
 513 sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
 514 {
 515         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 516         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 517         unsigned int next_f;
 518
 519         raw_spin_lock(&sg_policy->update_lock);
 520
 521         sugov_iowait_boost(sg_cpu, time, flags);
 522         sg_cpu->last_update = time;
 523
 524         ignore_dl_rate_limit(sg_cpu, sg_policy);
 525
 526         if (sugov_should_update_freq(sg_policy, time)) {
 527                 next_f = sugov_next_freq_shared(sg_cpu, time);
 528
 529                 if (sg_policy->policy->fast_switch_enabled)
 530                         sugov_fast_switch(sg_policy, time, next_f);
 531                 else
 532                         sugov_deferred_update(sg_policy, time, next_f);
 533         }
 534
 535         raw_spin_unlock(&sg_policy->update_lock);
 536 }
 537
 538 static void sugov_work(struct kthread_work *work)
 539 {
 540         struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
 541         unsigned int freq;
 542         unsigned long flags;
 543
 544         /*
 545          * Hold sg_policy->update_lock shortly to handle the case where:
 546          * incase sg_policy->next_freq is read here, and then updated by
 547          * sugov_deferred_update() just before work_in_progress is set to false
 548          * here, we may miss queueing the new update.
 549          *
 550          * Note: If a work was queued after the update_lock is released,
 551          * sugov_work() will just be called again by kthread_work code; and the
 552          * request will be proceed before the sugov thread sleeps.
 553          */
 554         raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
 555         freq = sg_policy->next_freq;
 556         sg_policy->work_in_progress = false;
 557         raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
 558
 559         mutex_lock(&sg_policy->work_lock);
 560         __cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
 561         mutex_unlock(&sg_policy->work_lock);
 562 }
 563
 564 static void sugov_irq_work(struct irq_work *irq_work)
 565 {
 566         struct sugov_policy *sg_policy;
 567
 568         sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
 569
 570         kthread_queue_work(&sg_policy->worker, &sg_policy->work);
 571 }
 572
 573 /************************** sysfs interface ************************/
 574
 575 static struct sugov_tunables *global_tunables;
 576 static DEFINE_MUTEX(global_tunables_lock);
 577
 578 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
 579 {
 580         return container_of(attr_set, struct sugov_tunables, attr_set);
 581 }
 582
 583 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
 584 {
 585         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 586
 587         return sprintf(buf, "%u\n", tunables->rate_limit_us);
 588 }
 589
 590 static ssize_t
 591 rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
 592 {
 593         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 594         struct sugov_policy *sg_policy;
 595         unsigned int rate_limit_us;
 596
 597         if (kstrtouint(buf, 10, &rate_limit_us))
 598                 return -EINVAL;
 599
 600         tunables->rate_limit_us = rate_limit_us;
 601
 602         list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
 603                 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
 604
 605         return count;
 606 }
 607
 608 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
 609
 610 static struct attribute *sugov_attrs[] = {
 611         &rate_limit_us.attr,
 612         NULL
 613 };
 614 ATTRIBUTE_GROUPS(sugov);
 615
 616 static struct kobj_type sugov_tunables_ktype = {
 617         .default_groups = sugov_groups,
 618         .sysfs_ops = &governor_sysfs_ops,
 619 };
 620
 621 /********************** cpufreq governor interface *********************/
 622
 623 struct cpufreq_governor schedutil_gov;
 624
 625 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
 626 {
 627         struct sugov_policy *sg_policy;
 628
 629         sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
 630         if (!sg_policy)
 631                 return NULL;
 632
 633         sg_policy->policy = policy;
 634         raw_spin_lock_init(&sg_policy->update_lock);
 635         return sg_policy;
 636 }
 637
 638 static void sugov_policy_free(struct sugov_policy *sg_policy)
 639 {
 640         kfree(sg_policy);
 641 }
 642
 643 static int sugov_kthread_create(struct sugov_policy *sg_policy)
 644 {
 645         struct task_struct *thread;
 646         struct sched_attr attr = {
 647                 .size           = sizeof(struct sched_attr),
 648                 .sched_policy   = SCHED_DEADLINE,
 649                 .sched_flags    = SCHED_FLAG_SUGOV,
 650                 .sched_nice     = 0,
 651                 .sched_priority = 0,
 652                 /*
 653                  * Fake (unused) bandwidth; workaround to "fix"
 654                  * priority inheritance.
 655                  */
 656                 .sched_runtime  =  1000000,
 657                 .sched_deadline = 10000000,
 658                 .sched_period   = 10000000,
 659         };
 660         struct cpufreq_policy *policy = sg_policy->policy;
 661         int ret;
 662
 663         /* kthread only required for slow path */
 664         if (policy->fast_switch_enabled)
 665                 return 0;
 666
 667         kthread_init_work(&sg_policy->work, sugov_work);
 668         kthread_init_worker(&sg_policy->worker);
 669         thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
 670                                 "sugov:%d",
 671                                 cpumask_first(policy->related_cpus));
 672         if (IS_ERR(thread)) {
 673                 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
 674                 return PTR_ERR(thread);
 675         }
 676
 677         ret = sched_setattr_nocheck(thread, &attr);
 678         if (ret) {
 679                 kthread_stop(thread);
 680                 pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
 681                 return ret;
 682         }
 683
 684         sg_policy->thread = thread;
 685         kthread_bind_mask(thread, policy->related_cpus);
 686         init_irq_work(&sg_policy->irq_work, sugov_irq_work);
 687         mutex_init(&sg_policy->work_lock);
 688
 689         wake_up_process(thread);
 690
 691         return 0;
 692 }
 693
 694 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
 695 {
 696         /* kthread only required for slow path */
 697         if (sg_policy->policy->fast_switch_enabled)
 698                 return;
 699
 700         kthread_flush_worker(&sg_policy->worker);
 701         kthread_stop(sg_policy->thread);
 702         mutex_destroy(&sg_policy->work_lock);
 703 }
 704
 705 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
 706 {
 707         struct sugov_tunables *tunables;
 708
 709         tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
 710         if (tunables) {
 711                 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
 712                 if (!have_governor_per_policy())
 713                         global_tunables = tunables;
 714         }
 715         return tunables;
 716 }
 717
 718 static void sugov_tunables_free(struct sugov_tunables *tunables)
 719 {
 720         if (!have_governor_per_policy())
 721                 global_tunables = NULL;
 722
 723         kfree(tunables);
 724 }
 725
 726 static int sugov_init(struct cpufreq_policy *policy)
 727 {
 728         struct sugov_policy *sg_policy;
 729         struct sugov_tunables *tunables;
 730         int ret = 0;
 731
 732         /* State should be equivalent to EXIT */
 733         if (policy->governor_data)
 734                 return -EBUSY;
 735
 736         cpufreq_enable_fast_switch(policy);
 737
 738         sg_policy = sugov_policy_alloc(policy);
 739         if (!sg_policy) {
 740                 ret = -ENOMEM;
 741                 goto disable_fast_switch;
 742         }
 743
 744         ret = sugov_kthread_create(sg_policy);
 745         if (ret)
 746                 goto free_sg_policy;
 747
 748         mutex_lock(&global_tunables_lock);
 749
 750         if (global_tunables) {
 751                 if (WARN_ON(have_governor_per_policy())) {
 752                         ret = -EINVAL;
 753                         goto stop_kthread;
 754                 }
 755                 policy->governor_data = sg_policy;
 756                 sg_policy->tunables = global_tunables;
 757
 758                 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
 759                 goto out;
 760         }
 761
 762         tunables = sugov_tunables_alloc(sg_policy);
 763         if (!tunables) {
 764                 ret = -ENOMEM;
 765                 goto stop_kthread;
 766         }
 767
 768         tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
 769
 770         policy->governor_data = sg_policy;
 771         sg_policy->tunables = tunables;
 772
 773         ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
 774                                    get_governor_parent_kobj(policy), "%s",
 775                                    schedutil_gov.name);
 776         if (ret)
 777                 goto fail;
 778
 779 out:
 780         mutex_unlock(&global_tunables_lock);
 781         return 0;
 782
 783 fail:
 784         kobject_put(&tunables->attr_set.kobj);
 785         policy->governor_data = NULL;
 786         sugov_tunables_free(tunables);
 787
 788 stop_kthread:
 789         sugov_kthread_stop(sg_policy);
 790         mutex_unlock(&global_tunables_lock);
 791
 792 free_sg_policy:
 793         sugov_policy_free(sg_policy);
 794
 795 disable_fast_switch:
 796         cpufreq_disable_fast_switch(policy);
 797
 798         pr_err("initialization failed (error %d)\n", ret);
 799         return ret;
 800 }
 801
 802 static void sugov_exit(struct cpufreq_policy *policy)
 803 {
 804         struct sugov_policy *sg_policy = policy->governor_data;
 805         struct sugov_tunables *tunables = sg_policy->tunables;
 806         unsigned int count;
 807
 808         mutex_lock(&global_tunables_lock);
 809
 810         count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
 811         policy->governor_data = NULL;
 812         if (!count)
 813                 sugov_tunables_free(tunables);
 814
 815         mutex_unlock(&global_tunables_lock);
 816
 817         sugov_kthread_stop(sg_policy);
 818         sugov_policy_free(sg_policy);
 819         cpufreq_disable_fast_switch(policy);
 820 }
 821
 822 static int sugov_start(struct cpufreq_policy *policy)
 823 {
 824         struct sugov_policy *sg_policy = policy->governor_data;
 825         unsigned int cpu;
 826
 827         sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
 828         sg_policy->last_freq_update_time        = 0;
 829         sg_policy->next_freq                    = 0;
 830         sg_policy->work_in_progress             = false;
 831         sg_policy->limits_changed               = false;
 832         sg_policy->need_freq_update             = false;
 833         sg_policy->cached_raw_freq              = 0;
 834
 835         for_each_cpu(cpu, policy->cpus) {
 836                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 837
 838                 memset(sg_cpu, 0, sizeof(*sg_cpu));
 839                 sg_cpu->cpu                     = cpu;
 840                 sg_cpu->sg_policy               = sg_policy;
 841         }
 842
 843         for_each_cpu(cpu, policy->cpus) {
 844                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 845
 846                 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
 847                                              policy_is_shared(policy) ?
 848                                                         sugov_update_shared :
 849                                                         sugov_update_single);
 850         }
 851         return 0;
 852 }
 853
 854 static void sugov_stop(struct cpufreq_policy *policy)
 855 {
 856         struct sugov_policy *sg_policy = policy->governor_data;
 857         unsigned int cpu;
 858
 859         for_each_cpu(cpu, policy->cpus)
 860                 cpufreq_remove_update_util_hook(cpu);
 861
 862         synchronize_rcu();
 863
 864         if (!policy->fast_switch_enabled) {
 865                 irq_work_sync(&sg_policy->irq_work);
 866                 kthread_cancel_work_sync(&sg_policy->work);
 867         }
 868 }
 869
 870 static void sugov_limits(struct cpufreq_policy *policy)
 871 {
 872         struct sugov_policy *sg_policy = policy->governor_data;
 873
 874         if (!policy->fast_switch_enabled) {
 875                 mutex_lock(&sg_policy->work_lock);
 876                 cpufreq_policy_apply_limits(policy);
 877                 mutex_unlock(&sg_policy->work_lock);
 878         }
 879
 880         sg_policy->limits_changed = true;
 881 }
 882
 883 struct cpufreq_governor schedutil_gov = {
 884         .name                   = "schedutil",
 885         .owner                  = THIS_MODULE,
 886         .dynamic_switching      = true,
 887         .init                   = sugov_init,
 888         .exit                   = sugov_exit,
 889         .start                  = sugov_start,
 890         .stop                   = sugov_stop,
 891         .limits                 = sugov_limits,
 892 };
 893
 894 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
 895 struct cpufreq_governor *cpufreq_default_governor(void)
 896 {
 897         return &schedutil_gov;
 898 }
 899 #endif
 900
 901 cpufreq_governor_init(schedutil_gov);
 902
 903 #ifdef CONFIG_ENERGY_MODEL
 904 extern bool sched_energy_update;
 905 extern struct mutex sched_energy_mutex;
 906
 907 static void rebuild_sd_workfn(struct work_struct *work)
 908 {
 909         mutex_lock(&sched_energy_mutex);
 910         sched_energy_update = true;
 911         rebuild_sched_domains();
 912         sched_energy_update = false;
 913         mutex_unlock(&sched_energy_mutex);
 914 }
 915 static DECLARE_WORK(rebuild_sd_work, rebuild_sd_workfn);
 916
 917 /*
 918  * EAS shouldn't be attempted without sugov, so rebuild the sched_domains
 919  * on governor changes to make sure the scheduler knows about it.
 920  */
 921 void sched_cpufreq_governor_change(struct cpufreq_policy *policy,
 922                                   struct cpufreq_governor *old_gov)
 923 {
 924         if (old_gov == &schedutil_gov || policy->governor == &schedutil_gov) {
 925                 /*
 926                  * When called from the cpufreq_register_driver() path, the
 927                  * cpu_hotplug_lock is already held, so use a work item to
 928                  * avoid nested locking in rebuild_sched_domains().
 929                  */
 930                 schedule_work(&rebuild_sd_work);
 931         }
 932
 933 }
 934 #endif