kernel/sched/cpufreq_schedutil.c

   1 /*
   2  * CPUFreq governor based on scheduler-provided CPU utilization data.
   3  *
   4  * Copyright (C) 2016, Intel Corporation
   5  * Author: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
   6  *
   7  * This program is free software; you can redistribute it and/or modify
   8  * it under the terms of the GNU General Public License version 2 as
   9  * published by the Free Software Foundation.
  10  */
  11
  12 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  13
  14 #include "sched.h"
  15
  16 #include <trace/events/power.h>
  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           min;
  58         unsigned long           max;
  59
  60         /* The field below is for single-CPU policies only: */
  61 #ifdef CONFIG_NO_HZ_COMMON
  62         unsigned long           saved_idle_calls;
  63 #endif
  64 };
  65
  66 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
  67
  68 /************************ Governor internals ***********************/
  69
  70 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
  71 {
  72         s64 delta_ns;
  73
  74         /*
  75          * Since cpufreq_update_util() is called with rq->lock held for
  76          * the @target_cpu, our per-CPU data is fully serialized.
  77          *
  78          * However, drivers cannot in general deal with cross-CPU
  79          * requests, so while get_next_freq() will work, our
  80          * sugov_update_commit() call may not for the fast switching platforms.
  81          *
  82          * Hence stop here for remote requests if they aren't supported
  83          * by the hardware, as calculating the frequency is pointless if
  84          * we cannot in fact act on it.
  85          *
  86          * This is needed on the slow switching platforms too to prevent CPUs
  87          * going offline from leaving stale IRQ work items behind.
  88          */
  89         if (!cpufreq_this_cpu_can_update(sg_policy->policy))
  90                 return false;
  91
  92         if (unlikely(sg_policy->limits_changed)) {
  93                 sg_policy->limits_changed = false;
  94                 sg_policy->need_freq_update = true;
  95                 return true;
  96         }
  97
  98         delta_ns = time - sg_policy->last_freq_update_time;
  99
 100         return delta_ns >= sg_policy->freq_update_delay_ns;
 101 }
 102
 103 static bool sugov_update_next_freq(struct sugov_policy *sg_policy, u64 time,
 104                                    unsigned int next_freq)
 105 {
 106         if (sg_policy->next_freq == next_freq)
 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         struct cpufreq_policy *policy = sg_policy->policy;
 119         int cpu;
 120
 121         if (!sugov_update_next_freq(sg_policy, time, next_freq))
 122                 return;
 123
 124         next_freq = cpufreq_driver_fast_switch(policy, next_freq);
 125         if (!next_freq)
 126                 return;
 127
 128         policy->cur = next_freq;
 129
 130         if (trace_cpu_frequency_enabled()) {
 131                 for_each_cpu(cpu, policy->cpus)
 132                         trace_cpu_frequency(next_freq, cpu);
 133         }
 134 }
 135
 136 static void sugov_deferred_update(struct sugov_policy *sg_policy, u64 time,
 137                                   unsigned int next_freq)
 138 {
 139         if (!sugov_update_next_freq(sg_policy, time, next_freq))
 140                 return;
 141
 142         if (!sg_policy->work_in_progress) {
 143                 sg_policy->work_in_progress = true;
 144                 irq_work_queue(&sg_policy->irq_work);
 145         }
 146 }
 147
 148 /**
 149  * get_next_freq - Compute a new frequency for a given cpufreq policy.
 150  * @sg_policy: schedutil policy object to compute the new frequency for.
 151  * @util: Current CPU utilization.
 152  * @max: CPU capacity.
 153  *
 154  * If the utilization is frequency-invariant, choose the new frequency to be
 155  * proportional to it, that is
 156  *
 157  * next_freq = C * max_freq * util / max
 158  *
 159  * Otherwise, approximate the would-be frequency-invariant utilization by
 160  * util_raw * (curr_freq / max_freq) which leads to
 161  *
 162  * next_freq = C * curr_freq * util_raw / max
 163  *
 164  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
 165  *
 166  * The lowest driver-supported frequency which is equal or greater than the raw
 167  * next_freq (as calculated above) is returned, subject to policy min/max and
 168  * cpufreq driver limitations.
 169  */
 170 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
 171                                   unsigned long util, unsigned long max)
 172 {
 173         struct cpufreq_policy *policy = sg_policy->policy;
 174         unsigned int freq = arch_scale_freq_invariant() ?
 175                                 policy->cpuinfo.max_freq : policy->cur;
 176
 177         freq = (freq + (freq >> 2)) * util / max;
 178
 179         if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
 180                 return sg_policy->next_freq;
 181
 182         sg_policy->need_freq_update = false;
 183         sg_policy->cached_raw_freq = freq;
 184         return cpufreq_driver_resolve_freq(policy, freq);
 185 }
 186
 187 /*
 188  * This function computes an effective utilization for the given CPU, to be
 189  * used for frequency selection given the linear relation: f = u * f_max.
 190  *
 191  * The scheduler tracks the following metrics:
 192  *
 193  *   cpu_util_{cfs,rt,dl,irq}()
 194  *   cpu_bw_dl()
 195  *
 196  * Where the cfs,rt and dl util numbers are tracked with the same metric and
 197  * synchronized windows and are thus directly comparable.
 198  *
 199  * The cfs,rt,dl utilization are the running times measured with rq->clock_task
 200  * which excludes things like IRQ and steal-time. These latter are then accrued
 201  * in the irq utilization.
 202  *
 203  * The DL bandwidth number otoh is not a measured metric but a value computed
 204  * based on the task model parameters and gives the minimal utilization
 205  * required to meet deadlines.
 206  */
 207 static unsigned long sugov_get_util(struct sugov_cpu *sg_cpu)
 208 {
 209         struct rq *rq = cpu_rq(sg_cpu->cpu);
 210         unsigned long util, irq, max;
 211
 212         sg_cpu->max = max = arch_scale_cpu_capacity(NULL, sg_cpu->cpu);
 213         sg_cpu->bw_dl = cpu_bw_dl(rq);
 214
 215         if (rt_rq_is_runnable(&rq->rt))
 216                 return max;
 217
 218         /*
 219          * Early check to see if IRQ/steal time saturates the CPU, can be
 220          * because of inaccuracies in how we track these -- see
 221          * update_irq_load_avg().
 222          */
 223         irq = cpu_util_irq(rq);
 224         if (unlikely(irq >= max))
 225                 return max;
 226
 227         /*
 228          * Because the time spend on RT/DL tasks is visible as 'lost' time to
 229          * CFS tasks and we use the same metric to track the effective
 230          * utilization (PELT windows are synchronized) we can directly add them
 231          * to obtain the CPU's actual utilization.
 232          */
 233         util = cpu_util_cfs(rq);
 234         util += cpu_util_rt(rq);
 235
 236         /*
 237          * We do not make cpu_util_dl() a permanent part of this sum because we
 238          * want to use cpu_bw_dl() later on, but we need to check if the
 239          * CFS+RT+DL sum is saturated (ie. no idle time) such that we select
 240          * f_max when there is no idle time.
 241          *
 242          * NOTE: numerical errors or stop class might cause us to not quite hit
 243          * saturation when we should -- something for later.
 244          */
 245         if ((util + cpu_util_dl(rq)) >= max)
 246                 return max;
 247
 248         /*
 249          * There is still idle time; further improve the number by using the
 250          * irq metric. Because IRQ/steal time is hidden from the task clock we
 251          * need to scale the task numbers:
 252          *
 253          *              1 - irq
 254          *   U' = irq + ------- * U
 255          *                max
 256          */
 257         util = scale_irq_capacity(util, irq, max);
 258         util += irq;
 259
 260         /*
 261          * Bandwidth required by DEADLINE must always be granted while, for
 262          * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
 263          * to gracefully reduce the frequency when no tasks show up for longer
 264          * periods of time.
 265          *
 266          * Ideally we would like to set bw_dl as min/guaranteed freq and util +
 267          * bw_dl as requested freq. However, cpufreq is not yet ready for such
 268          * an interface. So, we only do the latter for now.
 269          */
 270         return min(max, util + sg_cpu->bw_dl);
 271 }
 272
 273 /**
 274  * sugov_iowait_reset() - Reset the IO boost status of a CPU.
 275  * @sg_cpu: the sugov data for the CPU to boost
 276  * @time: the update time from the caller
 277  * @set_iowait_boost: true if an IO boost has been requested
 278  *
 279  * The IO wait boost of a task is disabled after a tick since the last update
 280  * of a CPU. If a new IO wait boost is requested after more then a tick, then
 281  * we enable the boost starting from the minimum frequency, which improves
 282  * energy efficiency by ignoring sporadic wakeups from IO.
 283  */
 284 static bool sugov_iowait_reset(struct sugov_cpu *sg_cpu, u64 time,
 285                                bool set_iowait_boost)
 286 {
 287         s64 delta_ns = time - sg_cpu->last_update;
 288
 289         /* Reset boost only if a tick has elapsed since last request */
 290         if (delta_ns <= TICK_NSEC)
 291                 return false;
 292
 293         sg_cpu->iowait_boost = set_iowait_boost ? sg_cpu->min : 0;
 294         sg_cpu->iowait_boost_pending = set_iowait_boost;
 295
 296         return true;
 297 }
 298
 299 /**
 300  * sugov_iowait_boost() - Updates the IO boost status of a CPU.
 301  * @sg_cpu: the sugov data for the CPU to boost
 302  * @time: the update time from the caller
 303  * @flags: SCHED_CPUFREQ_IOWAIT if the task is waking up after an IO wait
 304  *
 305  * Each time a task wakes up after an IO operation, the CPU utilization can be
 306  * boosted to a certain utilization which doubles at each "frequent and
 307  * successive" wakeup from IO, ranging from the utilization of the minimum
 308  * OPP to the utilization of the maximum OPP.
 309  * To keep doubling, an IO boost has to be requested at least once per tick,
 310  * otherwise we restart from the utilization of the minimum OPP.
 311  */
 312 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
 313                                unsigned int flags)
 314 {
 315         bool set_iowait_boost = flags & SCHED_CPUFREQ_IOWAIT;
 316
 317         /* Reset boost if the CPU appears to have been idle enough */
 318         if (sg_cpu->iowait_boost &&
 319             sugov_iowait_reset(sg_cpu, time, set_iowait_boost))
 320                 return;
 321
 322         /* Boost only tasks waking up after IO */
 323         if (!set_iowait_boost)
 324                 return;
 325
 326         /* Ensure boost doubles only one time at each request */
 327         if (sg_cpu->iowait_boost_pending)
 328                 return;
 329         sg_cpu->iowait_boost_pending = true;
 330
 331         /* Double the boost at each request */
 332         if (sg_cpu->iowait_boost) {
 333                 sg_cpu->iowait_boost =
 334                         min_t(unsigned int, sg_cpu->iowait_boost << 1, SCHED_CAPACITY_SCALE);
 335                 return;
 336         }
 337
 338         /* First wakeup after IO: start with minimum boost */
 339         sg_cpu->iowait_boost = sg_cpu->min;
 340 }
 341
 342 /**
 343  * sugov_iowait_apply() - Apply the IO boost to a CPU.
 344  * @sg_cpu: the sugov data for the cpu to boost
 345  * @time: the update time from the caller
 346  * @util: the utilization to (eventually) boost
 347  * @max: the maximum value the utilization can be boosted to
 348  *
 349  * A CPU running a task which woken up after an IO operation can have its
 350  * utilization boosted to speed up the completion of those IO operations.
 351  * The IO boost value is increased each time a task wakes up from IO, in
 352  * sugov_iowait_apply(), and it's instead decreased by this function,
 353  * each time an increase has not been requested (!iowait_boost_pending).
 354  *
 355  * A CPU which also appears to have been idle for at least one tick has also
 356  * its IO boost utilization reset.
 357  *
 358  * This mechanism is designed to boost high frequently IO waiting tasks, while
 359  * being more conservative on tasks which does sporadic IO operations.
 360  */
 361 static unsigned long sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
 362                                         unsigned long util, unsigned long max)
 363 {
 364         unsigned long boost;
 365
 366         /* No boost currently required */
 367         if (!sg_cpu->iowait_boost)
 368                 return util;
 369
 370         /* Reset boost if the CPU appears to have been idle enough */
 371         if (sugov_iowait_reset(sg_cpu, time, false))
 372                 return util;
 373
 374         if (!sg_cpu->iowait_boost_pending) {
 375                 /*
 376                  * No boost pending; reduce the boost value.
 377                  */
 378                 sg_cpu->iowait_boost >>= 1;
 379                 if (sg_cpu->iowait_boost < sg_cpu->min) {
 380                         sg_cpu->iowait_boost = 0;
 381                         return util;
 382                 }
 383         }
 384
 385         sg_cpu->iowait_boost_pending = false;
 386
 387         /*
 388          * @util is already in capacity scale; convert iowait_boost
 389          * into the same scale so we can compare.
 390          */
 391         boost = (sg_cpu->iowait_boost * max) >> SCHED_CAPACITY_SHIFT;
 392         return max(boost, util);
 393 }
 394
 395 #ifdef CONFIG_NO_HZ_COMMON
 396 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
 397 {
 398         unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
 399         bool ret = idle_calls == sg_cpu->saved_idle_calls;
 400
 401         sg_cpu->saved_idle_calls = idle_calls;
 402         return ret;
 403 }
 404 #else
 405 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
 406 #endif /* CONFIG_NO_HZ_COMMON */
 407
 408 /*
 409  * Make sugov_should_update_freq() ignore the rate limit when DL
 410  * has increased the utilization.
 411  */
 412 static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu, struct sugov_policy *sg_policy)
 413 {
 414         if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_dl)
 415                 sg_policy->limits_changed = true;
 416 }
 417
 418 static void sugov_update_single(struct update_util_data *hook, u64 time,
 419                                 unsigned int flags)
 420 {
 421         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 422         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 423         unsigned long util, max;
 424         unsigned int next_f;
 425         bool busy;
 426
 427         sugov_iowait_boost(sg_cpu, time, flags);
 428         sg_cpu->last_update = time;
 429
 430         ignore_dl_rate_limit(sg_cpu, sg_policy);
 431
 432         if (!sugov_should_update_freq(sg_policy, time))
 433                 return;
 434
 435         /* Limits may have changed, don't skip frequency update */
 436         busy = !sg_policy->need_freq_update && sugov_cpu_is_busy(sg_cpu);
 437
 438         util = sugov_get_util(sg_cpu);
 439         max = sg_cpu->max;
 440         util = sugov_iowait_apply(sg_cpu, time, util, max);
 441         next_f = get_next_freq(sg_policy, util, max);
 442         /*
 443          * Do not reduce the frequency if the CPU has not been idle
 444          * recently, as the reduction is likely to be premature then.
 445          */
 446         if (busy && next_f < sg_policy->next_freq) {
 447                 next_f = sg_policy->next_freq;
 448
 449                 /* Reset cached freq as next_freq has changed */
 450                 sg_policy->cached_raw_freq = 0;
 451         }
 452
 453         /*
 454          * This code runs under rq->lock for the target CPU, so it won't run
 455          * concurrently on two different CPUs for the same target and it is not
 456          * necessary to acquire the lock in the fast switch case.
 457          */
 458         if (sg_policy->policy->fast_switch_enabled) {
 459                 sugov_fast_switch(sg_policy, time, next_f);
 460         } else {
 461                 raw_spin_lock(&sg_policy->update_lock);
 462                 sugov_deferred_update(sg_policy, time, next_f);
 463                 raw_spin_unlock(&sg_policy->update_lock);
 464         }
 465 }
 466
 467 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
 468 {
 469         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 470         struct cpufreq_policy *policy = sg_policy->policy;
 471         unsigned long util = 0, max = 1;
 472         unsigned int j;
 473
 474         for_each_cpu(j, policy->cpus) {
 475                 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
 476                 unsigned long j_util, j_max;
 477
 478                 j_util = sugov_get_util(j_sg_cpu);
 479                 j_max = j_sg_cpu->max;
 480                 j_util = sugov_iowait_apply(j_sg_cpu, time, j_util, j_max);
 481
 482                 if (j_util * max > j_max * util) {
 483                         util = j_util;
 484                         max = j_max;
 485                 }
 486         }
 487
 488         return get_next_freq(sg_policy, util, max);
 489 }
 490
 491 static void
 492 sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
 493 {
 494         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 495         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 496         unsigned int next_f;
 497
 498         raw_spin_lock(&sg_policy->update_lock);
 499
 500         sugov_iowait_boost(sg_cpu, time, flags);
 501         sg_cpu->last_update = time;
 502
 503         ignore_dl_rate_limit(sg_cpu, sg_policy);
 504
 505         if (sugov_should_update_freq(sg_policy, time)) {
 506                 next_f = sugov_next_freq_shared(sg_cpu, time);
 507
 508                 if (sg_policy->policy->fast_switch_enabled)
 509                         sugov_fast_switch(sg_policy, time, next_f);
 510                 else
 511                         sugov_deferred_update(sg_policy, time, next_f);
 512         }
 513
 514         raw_spin_unlock(&sg_policy->update_lock);
 515 }
 516
 517 static void sugov_work(struct kthread_work *work)
 518 {
 519         struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
 520         unsigned int freq;
 521         unsigned long flags;
 522
 523         /*
 524          * Hold sg_policy->update_lock shortly to handle the case where:
 525          * incase sg_policy->next_freq is read here, and then updated by
 526          * sugov_deferred_update() just before work_in_progress is set to false
 527          * here, we may miss queueing the new update.
 528          *
 529          * Note: If a work was queued after the update_lock is released,
 530          * sugov_work() will just be called again by kthread_work code; and the
 531          * request will be proceed before the sugov thread sleeps.
 532          */
 533         raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
 534         freq = sg_policy->next_freq;
 535         sg_policy->work_in_progress = false;
 536         raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
 537
 538         mutex_lock(&sg_policy->work_lock);
 539         __cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
 540         mutex_unlock(&sg_policy->work_lock);
 541 }
 542
 543 static void sugov_irq_work(struct irq_work *irq_work)
 544 {
 545         struct sugov_policy *sg_policy;
 546
 547         sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
 548
 549         kthread_queue_work(&sg_policy->worker, &sg_policy->work);
 550 }
 551
 552 /************************** sysfs interface ************************/
 553
 554 static struct sugov_tunables *global_tunables;
 555 static DEFINE_MUTEX(global_tunables_lock);
 556
 557 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
 558 {
 559         return container_of(attr_set, struct sugov_tunables, attr_set);
 560 }
 561
 562 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
 563 {
 564         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 565
 566         return sprintf(buf, "%u\n", tunables->rate_limit_us);
 567 }
 568
 569 static ssize_t
 570 rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
 571 {
 572         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 573         struct sugov_policy *sg_policy;
 574         unsigned int rate_limit_us;
 575
 576         if (kstrtouint(buf, 10, &rate_limit_us))
 577                 return -EINVAL;
 578
 579         tunables->rate_limit_us = rate_limit_us;
 580
 581         list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
 582                 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
 583
 584         return count;
 585 }
 586
 587 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
 588
 589 static struct attribute *sugov_attributes[] = {
 590         &rate_limit_us.attr,
 591         NULL
 592 };
 593
 594 static void sugov_tunables_free(struct kobject *kobj)
 595 {
 596         struct gov_attr_set *attr_set = container_of(kobj, struct gov_attr_set, kobj);
 597
 598         kfree(to_sugov_tunables(attr_set));
 599 }
 600
 601 static struct kobj_type sugov_tunables_ktype = {
 602         .default_attrs = sugov_attributes,
 603         .sysfs_ops = &governor_sysfs_ops,
 604         .release = &sugov_tunables_free,
 605 };
 606
 607 /********************** cpufreq governor interface *********************/
 608
 609 static struct cpufreq_governor schedutil_gov;
 610
 611 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
 612 {
 613         struct sugov_policy *sg_policy;
 614
 615         sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
 616         if (!sg_policy)
 617                 return NULL;
 618
 619         sg_policy->policy = policy;
 620         raw_spin_lock_init(&sg_policy->update_lock);
 621         return sg_policy;
 622 }
 623
 624 static void sugov_policy_free(struct sugov_policy *sg_policy)
 625 {
 626         kfree(sg_policy);
 627 }
 628
 629 static int sugov_kthread_create(struct sugov_policy *sg_policy)
 630 {
 631         struct task_struct *thread;
 632         struct sched_attr attr = {
 633                 .size           = sizeof(struct sched_attr),
 634                 .sched_policy   = SCHED_DEADLINE,
 635                 .sched_flags    = SCHED_FLAG_SUGOV,
 636                 .sched_nice     = 0,
 637                 .sched_priority = 0,
 638                 /*
 639                  * Fake (unused) bandwidth; workaround to "fix"
 640                  * priority inheritance.
 641                  */
 642                 .sched_runtime  =  1000000,
 643                 .sched_deadline = 10000000,
 644                 .sched_period   = 10000000,
 645         };
 646         struct cpufreq_policy *policy = sg_policy->policy;
 647         int ret;
 648
 649         /* kthread only required for slow path */
 650         if (policy->fast_switch_enabled)
 651                 return 0;
 652
 653         kthread_init_work(&sg_policy->work, sugov_work);
 654         kthread_init_worker(&sg_policy->worker);
 655         thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
 656                                 "sugov:%d",
 657                                 cpumask_first(policy->related_cpus));
 658         if (IS_ERR(thread)) {
 659                 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
 660                 return PTR_ERR(thread);
 661         }
 662
 663         ret = sched_setattr_nocheck(thread, &attr);
 664         if (ret) {
 665                 kthread_stop(thread);
 666                 pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
 667                 return ret;
 668         }
 669
 670         sg_policy->thread = thread;
 671         kthread_bind_mask(thread, policy->related_cpus);
 672         init_irq_work(&sg_policy->irq_work, sugov_irq_work);
 673         mutex_init(&sg_policy->work_lock);
 674
 675         wake_up_process(thread);
 676
 677         return 0;
 678 }
 679
 680 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
 681 {
 682         /* kthread only required for slow path */
 683         if (sg_policy->policy->fast_switch_enabled)
 684                 return;
 685
 686         kthread_flush_worker(&sg_policy->worker);
 687         kthread_stop(sg_policy->thread);
 688         mutex_destroy(&sg_policy->work_lock);
 689 }
 690
 691 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
 692 {
 693         struct sugov_tunables *tunables;
 694
 695         tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
 696         if (tunables) {
 697                 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
 698                 if (!have_governor_per_policy())
 699                         global_tunables = tunables;
 700         }
 701         return tunables;
 702 }
 703
 704 static void sugov_clear_global_tunables(void)
 705 {
 706         if (!have_governor_per_policy())
 707                 global_tunables = NULL;
 708 }
 709
 710 static int sugov_init(struct cpufreq_policy *policy)
 711 {
 712         struct sugov_policy *sg_policy;
 713         struct sugov_tunables *tunables;
 714         int ret = 0;
 715
 716         /* State should be equivalent to EXIT */
 717         if (policy->governor_data)
 718                 return -EBUSY;
 719
 720         cpufreq_enable_fast_switch(policy);
 721
 722         sg_policy = sugov_policy_alloc(policy);
 723         if (!sg_policy) {
 724                 ret = -ENOMEM;
 725                 goto disable_fast_switch;
 726         }
 727
 728         ret = sugov_kthread_create(sg_policy);
 729         if (ret)
 730                 goto free_sg_policy;
 731
 732         mutex_lock(&global_tunables_lock);
 733
 734         if (global_tunables) {
 735                 if (WARN_ON(have_governor_per_policy())) {
 736                         ret = -EINVAL;
 737                         goto stop_kthread;
 738                 }
 739                 policy->governor_data = sg_policy;
 740                 sg_policy->tunables = global_tunables;
 741
 742                 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
 743                 goto out;
 744         }
 745
 746         tunables = sugov_tunables_alloc(sg_policy);
 747         if (!tunables) {
 748                 ret = -ENOMEM;
 749                 goto stop_kthread;
 750         }
 751
 752         tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
 753
 754         policy->governor_data = sg_policy;
 755         sg_policy->tunables = tunables;
 756
 757         ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
 758                                    get_governor_parent_kobj(policy), "%s",
 759                                    schedutil_gov.name);
 760         if (ret)
 761                 goto fail;
 762
 763 out:
 764         mutex_unlock(&global_tunables_lock);
 765         return 0;
 766
 767 fail:
 768         kobject_put(&tunables->attr_set.kobj);
 769         policy->governor_data = NULL;
 770         sugov_clear_global_tunables();
 771
 772 stop_kthread:
 773         sugov_kthread_stop(sg_policy);
 774         mutex_unlock(&global_tunables_lock);
 775
 776 free_sg_policy:
 777         sugov_policy_free(sg_policy);
 778
 779 disable_fast_switch:
 780         cpufreq_disable_fast_switch(policy);
 781
 782         pr_err("initialization failed (error %d)\n", ret);
 783         return ret;
 784 }
 785
 786 static void sugov_exit(struct cpufreq_policy *policy)
 787 {
 788         struct sugov_policy *sg_policy = policy->governor_data;
 789         struct sugov_tunables *tunables = sg_policy->tunables;
 790         unsigned int count;
 791
 792         mutex_lock(&global_tunables_lock);
 793
 794         count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
 795         policy->governor_data = NULL;
 796         if (!count)
 797                 sugov_clear_global_tunables();
 798
 799         mutex_unlock(&global_tunables_lock);
 800
 801         sugov_kthread_stop(sg_policy);
 802         sugov_policy_free(sg_policy);
 803         cpufreq_disable_fast_switch(policy);
 804 }
 805
 806 static int sugov_start(struct cpufreq_policy *policy)
 807 {
 808         struct sugov_policy *sg_policy = policy->governor_data;
 809         unsigned int cpu;
 810
 811         sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
 812         sg_policy->last_freq_update_time        = 0;
 813         sg_policy->next_freq                    = 0;
 814         sg_policy->work_in_progress             = false;
 815         sg_policy->limits_changed               = false;
 816         sg_policy->need_freq_update             = false;
 817         sg_policy->cached_raw_freq              = 0;
 818
 819         for_each_cpu(cpu, policy->cpus) {
 820                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 821
 822                 memset(sg_cpu, 0, sizeof(*sg_cpu));
 823                 sg_cpu->cpu                     = cpu;
 824                 sg_cpu->sg_policy               = sg_policy;
 825                 sg_cpu->min                     =
 826                         (SCHED_CAPACITY_SCALE * policy->cpuinfo.min_freq) /
 827                         policy->cpuinfo.max_freq;
 828         }
 829
 830         for_each_cpu(cpu, policy->cpus) {
 831                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 832
 833                 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
 834                                              policy_is_shared(policy) ?
 835                                                         sugov_update_shared :
 836                                                         sugov_update_single);
 837         }
 838         return 0;
 839 }
 840
 841 static void sugov_stop(struct cpufreq_policy *policy)
 842 {
 843         struct sugov_policy *sg_policy = policy->governor_data;
 844         unsigned int cpu;
 845
 846         for_each_cpu(cpu, policy->cpus)
 847                 cpufreq_remove_update_util_hook(cpu);
 848
 849         synchronize_sched();
 850
 851         if (!policy->fast_switch_enabled) {
 852                 irq_work_sync(&sg_policy->irq_work);
 853                 kthread_cancel_work_sync(&sg_policy->work);
 854         }
 855 }
 856
 857 static void sugov_limits(struct cpufreq_policy *policy)
 858 {
 859         struct sugov_policy *sg_policy = policy->governor_data;
 860
 861         if (!policy->fast_switch_enabled) {
 862                 mutex_lock(&sg_policy->work_lock);
 863                 cpufreq_policy_apply_limits(policy);
 864                 mutex_unlock(&sg_policy->work_lock);
 865         }
 866
 867         sg_policy->limits_changed = true;
 868 }
 869
 870 static struct cpufreq_governor schedutil_gov = {
 871         .name                   = "schedutil",
 872         .owner                  = THIS_MODULE,
 873         .dynamic_switching      = true,
 874         .init                   = sugov_init,
 875         .exit                   = sugov_exit,
 876         .start                  = sugov_start,
 877         .stop                   = sugov_stop,
 878         .limits                 = sugov_limits,
 879 };
 880
 881 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
 882 struct cpufreq_governor *cpufreq_default_governor(void)
 883 {
 884         return &schedutil_gov;
 885 }
 886 #endif
 887
 888 static int __init sugov_register(void)
 889 {
 890         return cpufreq_register_governor(&schedutil_gov);
 891 }
 892 fs_initcall(sugov_register);