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                    need_freq_update;
  44 };
  45
  46 struct sugov_cpu {
  47         struct update_util_data update_util;
  48         struct sugov_policy     *sg_policy;
  49         unsigned int            cpu;
  50
  51         bool                    iowait_boost_pending;
  52         unsigned int            iowait_boost;
  53         u64                     last_update;
  54
  55         unsigned long           bw_dl;
  56         unsigned long           max;
  57
  58         /* The field below is for single-CPU policies only: */
  59 #ifdef CONFIG_NO_HZ_COMMON
  60         unsigned long           saved_idle_calls;
  61 #endif
  62 };
  63
  64 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
  65
  66 /************************ Governor internals ***********************/
  67
  68 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
  69 {
  70         s64 delta_ns;
  71
  72         /*
  73          * Since cpufreq_update_util() is called with rq->lock held for
  74          * the @target_cpu, our per-CPU data is fully serialized.
  75          *
  76          * However, drivers cannot in general deal with cross-CPU
  77          * requests, so while get_next_freq() will work, our
  78          * sugov_update_commit() call may not for the fast switching platforms.
  79          *
  80          * Hence stop here for remote requests if they aren't supported
  81          * by the hardware, as calculating the frequency is pointless if
  82          * we cannot in fact act on it.
  83          *
  84          * For the slow switching platforms, the kthread is always scheduled on
  85          * the right set of CPUs and any CPU can find the next frequency and
  86          * schedule the kthread.
  87          */
  88         if (sg_policy->policy->fast_switch_enabled &&
  89             !cpufreq_this_cpu_can_update(sg_policy->policy))
  90                 return false;
  91
  92         if (unlikely(sg_policy->need_freq_update))
  93                 return true;
  94
  95         delta_ns = time - sg_policy->last_freq_update_time;
  96
  97         return delta_ns >= sg_policy->freq_update_delay_ns;
  98 }
  99
 100 static bool sugov_update_next_freq(struct sugov_policy *sg_policy, u64 time,
 101                                    unsigned int next_freq)
 102 {
 103         if (sg_policy->next_freq == next_freq)
 104                 return false;
 105
 106         sg_policy->next_freq = next_freq;
 107         sg_policy->last_freq_update_time = time;
 108
 109         return true;
 110 }
 111
 112 static void sugov_fast_switch(struct sugov_policy *sg_policy, u64 time,
 113                               unsigned int next_freq)
 114 {
 115         struct cpufreq_policy *policy = sg_policy->policy;
 116
 117         if (!sugov_update_next_freq(sg_policy, time, next_freq))
 118                 return;
 119
 120         next_freq = cpufreq_driver_fast_switch(policy, next_freq);
 121         if (!next_freq)
 122                 return;
 123
 124         policy->cur = next_freq;
 125         trace_cpu_frequency(next_freq, smp_processor_id());
 126 }
 127
 128 static void sugov_deferred_update(struct sugov_policy *sg_policy, u64 time,
 129                                   unsigned int next_freq)
 130 {
 131         if (!sugov_update_next_freq(sg_policy, time, next_freq))
 132                 return;
 133
 134         if (!sg_policy->work_in_progress) {
 135                 sg_policy->work_in_progress = true;
 136                 irq_work_queue(&sg_policy->irq_work);
 137         }
 138 }
 139
 140 /**
 141  * get_next_freq - Compute a new frequency for a given cpufreq policy.
 142  * @sg_policy: schedutil policy object to compute the new frequency for.
 143  * @util: Current CPU utilization.
 144  * @max: CPU capacity.
 145  *
 146  * If the utilization is frequency-invariant, choose the new frequency to be
 147  * proportional to it, that is
 148  *
 149  * next_freq = C * max_freq * util / max
 150  *
 151  * Otherwise, approximate the would-be frequency-invariant utilization by
 152  * util_raw * (curr_freq / max_freq) which leads to
 153  *
 154  * next_freq = C * curr_freq * util_raw / max
 155  *
 156  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
 157  *
 158  * The lowest driver-supported frequency which is equal or greater than the raw
 159  * next_freq (as calculated above) is returned, subject to policy min/max and
 160  * cpufreq driver limitations.
 161  */
 162 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
 163                                   unsigned long util, unsigned long max)
 164 {
 165         struct cpufreq_policy *policy = sg_policy->policy;
 166         unsigned int freq = arch_scale_freq_invariant() ?
 167                                 policy->cpuinfo.max_freq : policy->cur;
 168
 169         freq = map_util_freq(util, freq, max);
 170
 171         if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
 172                 return sg_policy->next_freq;
 173
 174         sg_policy->need_freq_update = false;
 175         sg_policy->cached_raw_freq = freq;
 176         return cpufreq_driver_resolve_freq(policy, freq);
 177 }
 178
 179 /*
 180  * This function computes an effective utilization for the given CPU, to be
 181  * used for frequency selection given the linear relation: f = u * f_max.
 182  *
 183  * The scheduler tracks the following metrics:
 184  *
 185  *   cpu_util_{cfs,rt,dl,irq}()
 186  *   cpu_bw_dl()
 187  *
 188  * Where the cfs,rt and dl util numbers are tracked with the same metric and
 189  * synchronized windows and are thus directly comparable.
 190  *
 191  * The cfs,rt,dl utilization are the running times measured with rq->clock_task
 192  * which excludes things like IRQ and steal-time. These latter are then accrued
 193  * in the irq utilization.
 194  *
 195  * The DL bandwidth number otoh is not a measured metric but a value computed
 196  * based on the task model parameters and gives the minimal utilization
 197  * required to meet deadlines.
 198  */
 199 unsigned long schedutil_cpu_util(int cpu, unsigned long util_cfs,
 200                                  unsigned long max, enum schedutil_type type,
 201                                  struct task_struct *p)
 202 {
 203         unsigned long dl_util, util, irq;
 204         struct rq *rq = cpu_rq(cpu);
 205
 206         if (!IS_BUILTIN(CONFIG_UCLAMP_TASK) &&
 207             type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
 208                 return max;
 209         }
 210
 211         /*
 212          * Early check to see if IRQ/steal time saturates the CPU, can be
 213          * because of inaccuracies in how we track these -- see
 214          * update_irq_load_avg().
 215          */
 216         irq = cpu_util_irq(rq);
 217         if (unlikely(irq >= max))
 218                 return max;
 219
 220         /*
 221          * Because the time spend on RT/DL tasks is visible as 'lost' time to
 222          * CFS tasks and we use the same metric to track the effective
 223          * utilization (PELT windows are synchronized) we can directly add them
 224          * to obtain the CPU's actual utilization.
 225          *
 226          * CFS and RT utilization can be boosted or capped, depending on
 227          * utilization clamp constraints requested by currently RUNNABLE
 228          * tasks.
 229          * When there are no CFS RUNNABLE tasks, clamps are released and
 230          * frequency will be gracefully reduced with the utilization decay.
 231          */
 232         util = util_cfs + cpu_util_rt(rq);
 233         if (type == FREQUENCY_UTIL)
 234                 util = uclamp_util_with(rq, util, p);
 235
 236         dl_util = cpu_util_dl(rq);
 237
 238         /*
 239          * For frequency selection we do not make cpu_util_dl() a permanent part
 240          * of this sum because we want to use cpu_bw_dl() later on, but we need
 241          * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
 242          * that we select f_max when there is no idle time.
 243          *
 244          * NOTE: numerical errors or stop class might cause us to not quite hit
 245          * saturation when we should -- something for later.
 246          */
 247         if (util + dl_util >= max)
 248                 return max;
 249
 250         /*
 251          * OTOH, for energy computation we need the estimated running time, so
 252          * include util_dl and ignore dl_bw.
 253          */
 254         if (type == ENERGY_UTIL)
 255                 util += dl_util;
 256
 257         /*
 258          * There is still idle time; further improve the number by using the
 259          * irq metric. Because IRQ/steal time is hidden from the task clock we
 260          * need to scale the task numbers:
 261          *
 262          *              1 - irq
 263          *   U' = irq + ------- * U
 264          *                max
 265          */
 266         util = scale_irq_capacity(util, irq, max);
 267         util += irq;
 268
 269         /*
 270          * Bandwidth required by DEADLINE must always be granted while, for
 271          * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
 272          * to gracefully reduce the frequency when no tasks show up for longer
 273          * periods of time.
 274          *
 275          * Ideally we would like to set bw_dl as min/guaranteed freq and util +
 276          * bw_dl as requested freq. However, cpufreq is not yet ready for such
 277          * an interface. So, we only do the latter for now.
 278          */
 279         if (type == FREQUENCY_UTIL)
 280                 util += cpu_bw_dl(rq);
 281
 282         return min(max, util);
 283 }
 284
 285 static unsigned long sugov_get_util(struct sugov_cpu *sg_cpu)
 286 {
 287         struct rq *rq = cpu_rq(sg_cpu->cpu);
 288         unsigned long util = cpu_util_cfs(rq);
 289         unsigned long max = arch_scale_cpu_capacity(sg_cpu->cpu);
 290
 291         sg_cpu->max = max;
 292         sg_cpu->bw_dl = cpu_bw_dl(rq);
 293
 294         return schedutil_cpu_util(sg_cpu->cpu, util, max, FREQUENCY_UTIL, NULL);
 295 }
 296
 297 /**
 298  * sugov_iowait_reset() - Reset the IO boost status of a CPU.
 299  * @sg_cpu: the sugov data for the CPU to boost
 300  * @time: the update time from the caller
 301  * @set_iowait_boost: true if an IO boost has been requested
 302  *
 303  * The IO wait boost of a task is disabled after a tick since the last update
 304  * of a CPU. If a new IO wait boost is requested after more then a tick, then
 305  * we enable the boost starting from IOWAIT_BOOST_MIN, which improves energy
 306  * efficiency by ignoring sporadic wakeups from IO.
 307  */
 308 static bool sugov_iowait_reset(struct sugov_cpu *sg_cpu, u64 time,
 309                                bool set_iowait_boost)
 310 {
 311         s64 delta_ns = time - sg_cpu->last_update;
 312
 313         /* Reset boost only if a tick has elapsed since last request */
 314         if (delta_ns <= TICK_NSEC)
 315                 return false;
 316
 317         sg_cpu->iowait_boost = set_iowait_boost ? IOWAIT_BOOST_MIN : 0;
 318         sg_cpu->iowait_boost_pending = set_iowait_boost;
 319
 320         return true;
 321 }
 322
 323 /**
 324  * sugov_iowait_boost() - Updates the IO boost status of a CPU.
 325  * @sg_cpu: the sugov data for the CPU to boost
 326  * @time: the update time from the caller
 327  * @flags: SCHED_CPUFREQ_IOWAIT if the task is waking up after an IO wait
 328  *
 329  * Each time a task wakes up after an IO operation, the CPU utilization can be
 330  * boosted to a certain utilization which doubles at each "frequent and
 331  * successive" wakeup from IO, ranging from IOWAIT_BOOST_MIN to the utilization
 332  * of the maximum OPP.
 333  *
 334  * To keep doubling, an IO boost has to be requested at least once per tick,
 335  * otherwise we restart from the utilization of the minimum OPP.
 336  */
 337 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
 338                                unsigned int flags)
 339 {
 340         bool set_iowait_boost = flags & SCHED_CPUFREQ_IOWAIT;
 341
 342         /* Reset boost if the CPU appears to have been idle enough */
 343         if (sg_cpu->iowait_boost &&
 344             sugov_iowait_reset(sg_cpu, time, set_iowait_boost))
 345                 return;
 346
 347         /* Boost only tasks waking up after IO */
 348         if (!set_iowait_boost)
 349                 return;
 350
 351         /* Ensure boost doubles only one time at each request */
 352         if (sg_cpu->iowait_boost_pending)
 353                 return;
 354         sg_cpu->iowait_boost_pending = true;
 355
 356         /* Double the boost at each request */
 357         if (sg_cpu->iowait_boost) {
 358                 sg_cpu->iowait_boost =
 359                         min_t(unsigned int, sg_cpu->iowait_boost << 1, SCHED_CAPACITY_SCALE);
 360                 return;
 361         }
 362
 363         /* First wakeup after IO: start with minimum boost */
 364         sg_cpu->iowait_boost = IOWAIT_BOOST_MIN;
 365 }
 366
 367 /**
 368  * sugov_iowait_apply() - Apply the IO boost to a CPU.
 369  * @sg_cpu: the sugov data for the cpu to boost
 370  * @time: the update time from the caller
 371  * @util: the utilization to (eventually) boost
 372  * @max: the maximum value the utilization can be boosted to
 373  *
 374  * A CPU running a task which woken up after an IO operation can have its
 375  * utilization boosted to speed up the completion of those IO operations.
 376  * The IO boost value is increased each time a task wakes up from IO, in
 377  * sugov_iowait_apply(), and it's instead decreased by this function,
 378  * each time an increase has not been requested (!iowait_boost_pending).
 379  *
 380  * A CPU which also appears to have been idle for at least one tick has also
 381  * its IO boost utilization reset.
 382  *
 383  * This mechanism is designed to boost high frequently IO waiting tasks, while
 384  * being more conservative on tasks which does sporadic IO operations.
 385  */
 386 static unsigned long sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
 387                                         unsigned long util, unsigned long max)
 388 {
 389         unsigned long boost;
 390
 391         /* No boost currently required */
 392         if (!sg_cpu->iowait_boost)
 393                 return util;
 394
 395         /* Reset boost if the CPU appears to have been idle enough */
 396         if (sugov_iowait_reset(sg_cpu, time, false))
 397                 return util;
 398
 399         if (!sg_cpu->iowait_boost_pending) {
 400                 /*
 401                  * No boost pending; reduce the boost value.
 402                  */
 403                 sg_cpu->iowait_boost >>= 1;
 404                 if (sg_cpu->iowait_boost < IOWAIT_BOOST_MIN) {
 405                         sg_cpu->iowait_boost = 0;
 406                         return util;
 407                 }
 408         }
 409
 410         sg_cpu->iowait_boost_pending = false;
 411
 412         /*
 413          * @util is already in capacity scale; convert iowait_boost
 414          * into the same scale so we can compare.
 415          */
 416         boost = (sg_cpu->iowait_boost * max) >> SCHED_CAPACITY_SHIFT;
 417         return max(boost, util);
 418 }
 419
 420 #ifdef CONFIG_NO_HZ_COMMON
 421 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
 422 {
 423         unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
 424         bool ret = idle_calls == sg_cpu->saved_idle_calls;
 425
 426         sg_cpu->saved_idle_calls = idle_calls;
 427         return ret;
 428 }
 429 #else
 430 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
 431 #endif /* CONFIG_NO_HZ_COMMON */
 432
 433 /*
 434  * Make sugov_should_update_freq() ignore the rate limit when DL
 435  * has increased the utilization.
 436  */
 437 static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu, struct sugov_policy *sg_policy)
 438 {
 439         if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_dl)
 440                 sg_policy->need_freq_update = true;
 441 }
 442
 443 static void sugov_update_single(struct update_util_data *hook, u64 time,
 444                                 unsigned int flags)
 445 {
 446         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 447         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 448         unsigned long util, max;
 449         unsigned int next_f;
 450         bool busy;
 451
 452         sugov_iowait_boost(sg_cpu, time, flags);
 453         sg_cpu->last_update = time;
 454
 455         ignore_dl_rate_limit(sg_cpu, sg_policy);
 456
 457         if (!sugov_should_update_freq(sg_policy, time))
 458                 return;
 459
 460         busy = sugov_cpu_is_busy(sg_cpu);
 461
 462         util = sugov_get_util(sg_cpu);
 463         max = sg_cpu->max;
 464         util = sugov_iowait_apply(sg_cpu, time, util, max);
 465         next_f = get_next_freq(sg_policy, util, max);
 466         /*
 467          * Do not reduce the frequency if the CPU has not been idle
 468          * recently, as the reduction is likely to be premature then.
 469          */
 470         if (busy && next_f < sg_policy->next_freq) {
 471                 next_f = sg_policy->next_freq;
 472
 473                 /* Reset cached freq as next_freq has changed */
 474                 sg_policy->cached_raw_freq = 0;
 475         }
 476
 477         /*
 478          * This code runs under rq->lock for the target CPU, so it won't run
 479          * concurrently on two different CPUs for the same target and it is not
 480          * necessary to acquire the lock in the fast switch case.
 481          */
 482         if (sg_policy->policy->fast_switch_enabled) {
 483                 sugov_fast_switch(sg_policy, time, next_f);
 484         } else {
 485                 raw_spin_lock(&sg_policy->update_lock);
 486                 sugov_deferred_update(sg_policy, time, next_f);
 487                 raw_spin_unlock(&sg_policy->update_lock);
 488         }
 489 }
 490
 491 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
 492 {
 493         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 494         struct cpufreq_policy *policy = sg_policy->policy;
 495         unsigned long util = 0, max = 1;
 496         unsigned int j;
 497
 498         for_each_cpu(j, policy->cpus) {
 499                 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
 500                 unsigned long j_util, j_max;
 501
 502                 j_util = sugov_get_util(j_sg_cpu);
 503                 j_max = j_sg_cpu->max;
 504                 j_util = sugov_iowait_apply(j_sg_cpu, time, j_util, j_max);
 505
 506                 if (j_util * max > j_max * util) {
 507                         util = j_util;
 508                         max = j_max;
 509                 }
 510         }
 511
 512         return get_next_freq(sg_policy, util, max);
 513 }
 514
 515 static void
 516 sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
 517 {
 518         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 519         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 520         unsigned int next_f;
 521
 522         raw_spin_lock(&sg_policy->update_lock);
 523
 524         sugov_iowait_boost(sg_cpu, time, flags);
 525         sg_cpu->last_update = time;
 526
 527         ignore_dl_rate_limit(sg_cpu, sg_policy);
 528
 529         if (sugov_should_update_freq(sg_policy, time)) {
 530                 next_f = sugov_next_freq_shared(sg_cpu, time);
 531
 532                 if (sg_policy->policy->fast_switch_enabled)
 533                         sugov_fast_switch(sg_policy, time, next_f);
 534                 else
 535                         sugov_deferred_update(sg_policy, time, next_f);
 536         }
 537
 538         raw_spin_unlock(&sg_policy->update_lock);
 539 }
 540
 541 static void sugov_work(struct kthread_work *work)
 542 {
 543         struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
 544         unsigned int freq;
 545         unsigned long flags;
 546
 547         /*
 548          * Hold sg_policy->update_lock shortly to handle the case where:
 549          * incase sg_policy->next_freq is read here, and then updated by
 550          * sugov_deferred_update() just before work_in_progress is set to false
 551          * here, we may miss queueing the new update.
 552          *
 553          * Note: If a work was queued after the update_lock is released,
 554          * sugov_work() will just be called again by kthread_work code; and the
 555          * request will be proceed before the sugov thread sleeps.
 556          */
 557         raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
 558         freq = sg_policy->next_freq;
 559         sg_policy->work_in_progress = false;
 560         raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
 561
 562         mutex_lock(&sg_policy->work_lock);
 563         __cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
 564         mutex_unlock(&sg_policy->work_lock);
 565 }
 566
 567 static void sugov_irq_work(struct irq_work *irq_work)
 568 {
 569         struct sugov_policy *sg_policy;
 570
 571         sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
 572
 573         kthread_queue_work(&sg_policy->worker, &sg_policy->work);
 574 }
 575
 576 /************************** sysfs interface ************************/
 577
 578 static struct sugov_tunables *global_tunables;
 579 static DEFINE_MUTEX(global_tunables_lock);
 580
 581 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
 582 {
 583         return container_of(attr_set, struct sugov_tunables, attr_set);
 584 }
 585
 586 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
 587 {
 588         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 589
 590         return sprintf(buf, "%u\n", tunables->rate_limit_us);
 591 }
 592
 593 static ssize_t
 594 rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
 595 {
 596         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 597         struct sugov_policy *sg_policy;
 598         unsigned int rate_limit_us;
 599
 600         if (kstrtouint(buf, 10, &rate_limit_us))
 601                 return -EINVAL;
 602
 603         tunables->rate_limit_us = rate_limit_us;
 604
 605         list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
 606                 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
 607
 608         return count;
 609 }
 610
 611 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
 612
 613 static struct attribute *sugov_attrs[] = {
 614         &rate_limit_us.attr,
 615         NULL
 616 };
 617 ATTRIBUTE_GROUPS(sugov);
 618
 619 static struct kobj_type sugov_tunables_ktype = {
 620         .default_groups = sugov_groups,
 621         .sysfs_ops = &governor_sysfs_ops,
 622 };
 623
 624 /********************** cpufreq governor interface *********************/
 625
 626 struct cpufreq_governor schedutil_gov;
 627
 628 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
 629 {
 630         struct sugov_policy *sg_policy;
 631
 632         sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
 633         if (!sg_policy)
 634                 return NULL;
 635
 636         sg_policy->policy = policy;
 637         raw_spin_lock_init(&sg_policy->update_lock);
 638         return sg_policy;
 639 }
 640
 641 static void sugov_policy_free(struct sugov_policy *sg_policy)
 642 {
 643         kfree(sg_policy);
 644 }
 645
 646 static int sugov_kthread_create(struct sugov_policy *sg_policy)
 647 {
 648         struct task_struct *thread;
 649         struct sched_attr attr = {
 650                 .size           = sizeof(struct sched_attr),
 651                 .sched_policy   = SCHED_DEADLINE,
 652                 .sched_flags    = SCHED_FLAG_SUGOV,
 653                 .sched_nice     = 0,
 654                 .sched_priority = 0,
 655                 /*
 656                  * Fake (unused) bandwidth; workaround to "fix"
 657                  * priority inheritance.
 658                  */
 659                 .sched_runtime  =  1000000,
 660                 .sched_deadline = 10000000,
 661                 .sched_period   = 10000000,
 662         };
 663         struct cpufreq_policy *policy = sg_policy->policy;
 664         int ret;
 665
 666         /* kthread only required for slow path */
 667         if (policy->fast_switch_enabled)
 668                 return 0;
 669
 670         kthread_init_work(&sg_policy->work, sugov_work);
 671         kthread_init_worker(&sg_policy->worker);
 672         thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
 673                                 "sugov:%d",
 674                                 cpumask_first(policy->related_cpus));
 675         if (IS_ERR(thread)) {
 676                 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
 677                 return PTR_ERR(thread);
 678         }
 679
 680         ret = sched_setattr_nocheck(thread, &attr);
 681         if (ret) {
 682                 kthread_stop(thread);
 683                 pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
 684                 return ret;
 685         }
 686
 687         sg_policy->thread = thread;
 688         kthread_bind_mask(thread, policy->related_cpus);
 689         init_irq_work(&sg_policy->irq_work, sugov_irq_work);
 690         mutex_init(&sg_policy->work_lock);
 691
 692         wake_up_process(thread);
 693
 694         return 0;
 695 }
 696
 697 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
 698 {
 699         /* kthread only required for slow path */
 700         if (sg_policy->policy->fast_switch_enabled)
 701                 return;
 702
 703         kthread_flush_worker(&sg_policy->worker);
 704         kthread_stop(sg_policy->thread);
 705         mutex_destroy(&sg_policy->work_lock);
 706 }
 707
 708 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
 709 {
 710         struct sugov_tunables *tunables;
 711
 712         tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
 713         if (tunables) {
 714                 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
 715                 if (!have_governor_per_policy())
 716                         global_tunables = tunables;
 717         }
 718         return tunables;
 719 }
 720
 721 static void sugov_tunables_free(struct sugov_tunables *tunables)
 722 {
 723         if (!have_governor_per_policy())
 724                 global_tunables = NULL;
 725
 726         kfree(tunables);
 727 }
 728
 729 static int sugov_init(struct cpufreq_policy *policy)
 730 {
 731         struct sugov_policy *sg_policy;
 732         struct sugov_tunables *tunables;
 733         int ret = 0;
 734
 735         /* State should be equivalent to EXIT */
 736         if (policy->governor_data)
 737                 return -EBUSY;
 738
 739         cpufreq_enable_fast_switch(policy);
 740
 741         sg_policy = sugov_policy_alloc(policy);
 742         if (!sg_policy) {
 743                 ret = -ENOMEM;
 744                 goto disable_fast_switch;
 745         }
 746
 747         ret = sugov_kthread_create(sg_policy);
 748         if (ret)
 749                 goto free_sg_policy;
 750
 751         mutex_lock(&global_tunables_lock);
 752
 753         if (global_tunables) {
 754                 if (WARN_ON(have_governor_per_policy())) {
 755                         ret = -EINVAL;
 756                         goto stop_kthread;
 757                 }
 758                 policy->governor_data = sg_policy;
 759                 sg_policy->tunables = global_tunables;
 760
 761                 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
 762                 goto out;
 763         }
 764
 765         tunables = sugov_tunables_alloc(sg_policy);
 766         if (!tunables) {
 767                 ret = -ENOMEM;
 768                 goto stop_kthread;
 769         }
 770
 771         tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
 772
 773         policy->governor_data = sg_policy;
 774         sg_policy->tunables = tunables;
 775
 776         ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
 777                                    get_governor_parent_kobj(policy), "%s",
 778                                    schedutil_gov.name);
 779         if (ret)
 780                 goto fail;
 781
 782 out:
 783         mutex_unlock(&global_tunables_lock);
 784         return 0;
 785
 786 fail:
 787         kobject_put(&tunables->attr_set.kobj);
 788         policy->governor_data = NULL;
 789         sugov_tunables_free(tunables);
 790
 791 stop_kthread:
 792         sugov_kthread_stop(sg_policy);
 793         mutex_unlock(&global_tunables_lock);
 794
 795 free_sg_policy:
 796         sugov_policy_free(sg_policy);
 797
 798 disable_fast_switch:
 799         cpufreq_disable_fast_switch(policy);
 800
 801         pr_err("initialization failed (error %d)\n", ret);
 802         return ret;
 803 }
 804
 805 static void sugov_exit(struct cpufreq_policy *policy)
 806 {
 807         struct sugov_policy *sg_policy = policy->governor_data;
 808         struct sugov_tunables *tunables = sg_policy->tunables;
 809         unsigned int count;
 810
 811         mutex_lock(&global_tunables_lock);
 812
 813         count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
 814         policy->governor_data = NULL;
 815         if (!count)
 816                 sugov_tunables_free(tunables);
 817
 818         mutex_unlock(&global_tunables_lock);
 819
 820         sugov_kthread_stop(sg_policy);
 821         sugov_policy_free(sg_policy);
 822         cpufreq_disable_fast_switch(policy);
 823 }
 824
 825 static int sugov_start(struct cpufreq_policy *policy)
 826 {
 827         struct sugov_policy *sg_policy = policy->governor_data;
 828         unsigned int cpu;
 829
 830         sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
 831         sg_policy->last_freq_update_time        = 0;
 832         sg_policy->next_freq                    = 0;
 833         sg_policy->work_in_progress             = false;
 834         sg_policy->need_freq_update             = false;
 835         sg_policy->cached_raw_freq              = 0;
 836
 837         for_each_cpu(cpu, policy->cpus) {
 838                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 839
 840                 memset(sg_cpu, 0, sizeof(*sg_cpu));
 841                 sg_cpu->cpu                     = cpu;
 842                 sg_cpu->sg_policy               = sg_policy;
 843         }
 844
 845         for_each_cpu(cpu, policy->cpus) {
 846                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 847
 848                 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
 849                                              policy_is_shared(policy) ?
 850                                                         sugov_update_shared :
 851                                                         sugov_update_single);
 852         }
 853         return 0;
 854 }
 855
 856 static void sugov_stop(struct cpufreq_policy *policy)
 857 {
 858         struct sugov_policy *sg_policy = policy->governor_data;
 859         unsigned int cpu;
 860
 861         for_each_cpu(cpu, policy->cpus)
 862                 cpufreq_remove_update_util_hook(cpu);
 863
 864         synchronize_rcu();
 865
 866         if (!policy->fast_switch_enabled) {
 867                 irq_work_sync(&sg_policy->irq_work);
 868                 kthread_cancel_work_sync(&sg_policy->work);
 869         }
 870 }
 871
 872 static void sugov_limits(struct cpufreq_policy *policy)
 873 {
 874         struct sugov_policy *sg_policy = policy->governor_data;
 875
 876         if (!policy->fast_switch_enabled) {
 877                 mutex_lock(&sg_policy->work_lock);
 878                 cpufreq_policy_apply_limits(policy);
 879                 mutex_unlock(&sg_policy->work_lock);
 880         }
 881
 882         sg_policy->need_freq_update = true;
 883 }
 884
 885 struct cpufreq_governor schedutil_gov = {
 886         .name                   = "schedutil",
 887         .owner                  = THIS_MODULE,
 888         .dynamic_switching      = true,
 889         .init                   = sugov_init,
 890         .exit                   = sugov_exit,
 891         .start                  = sugov_start,
 892         .stop                   = sugov_stop,
 893         .limits                 = sugov_limits,
 894 };
 895
 896 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
 897 struct cpufreq_governor *cpufreq_default_governor(void)
 898 {
 899         return &schedutil_gov;
 900 }
 901 #endif
 902
 903 static int __init sugov_register(void)
 904 {
 905         return cpufreq_register_governor(&schedutil_gov);
 906 }
 907 fs_initcall(sugov_register);
 908
 909 #ifdef CONFIG_ENERGY_MODEL
 910 extern bool sched_energy_update;
 911 extern struct mutex sched_energy_mutex;
 912
 913 static void rebuild_sd_workfn(struct work_struct *work)
 914 {
 915         mutex_lock(&sched_energy_mutex);
 916         sched_energy_update = true;
 917         rebuild_sched_domains();
 918         sched_energy_update = false;
 919         mutex_unlock(&sched_energy_mutex);
 920 }
 921 static DECLARE_WORK(rebuild_sd_work, rebuild_sd_workfn);
 922
 923 /*
 924  * EAS shouldn't be attempted without sugov, so rebuild the sched_domains
 925  * on governor changes to make sure the scheduler knows about it.
 926  */
 927 void sched_cpufreq_governor_change(struct cpufreq_policy *policy,
 928                                   struct cpufreq_governor *old_gov)
 929 {
 930         if (old_gov == &schedutil_gov || policy->governor == &schedutil_gov) {
 931                 /*
 932                  * When called from the cpufreq_register_driver() path, the
 933                  * cpu_hotplug_lock is already held, so use a work item to
 934                  * avoid nested locking in rebuild_sched_domains().
 935                  */
 936                 schedule_work(&rebuild_sd_work);
 937         }
 938
 939 }
 940 #endif