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