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;
  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           util;
  57         unsigned long           bw_dl;
  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->need_freq_update)
 107                 sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
 108         else if (sg_policy->next_freq == next_freq)
 109                 return false;
 110
 111         sg_policy->next_freq = next_freq;
 112         sg_policy->last_freq_update_time = time;
 113
 114         return true;
 115 }
 116
 117 static void sugov_fast_switch(struct sugov_policy *sg_policy, u64 time,
 118                               unsigned int next_freq)
 119 {
 120         if (sugov_update_next_freq(sg_policy, time, next_freq))
 121                 cpufreq_driver_fast_switch(sg_policy->policy, next_freq);
 122 }
 123
 124 static void sugov_deferred_update(struct sugov_policy *sg_policy, u64 time,
 125                                   unsigned int next_freq)
 126 {
 127         if (!sugov_update_next_freq(sg_policy, time, next_freq))
 128                 return;
 129
 130         if (!sg_policy->work_in_progress) {
 131                 sg_policy->work_in_progress = true;
 132                 irq_work_queue(&sg_policy->irq_work);
 133         }
 134 }
 135
 136 /**
 137  * get_next_freq - Compute a new frequency for a given cpufreq policy.
 138  * @sg_policy: schedutil policy object to compute the new frequency for.
 139  * @util: Current CPU utilization.
 140  * @max: CPU capacity.
 141  *
 142  * If the utilization is frequency-invariant, choose the new frequency to be
 143  * proportional to it, that is
 144  *
 145  * next_freq = C * max_freq * util / max
 146  *
 147  * Otherwise, approximate the would-be frequency-invariant utilization by
 148  * util_raw * (curr_freq / max_freq) which leads to
 149  *
 150  * next_freq = C * curr_freq * util_raw / max
 151  *
 152  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
 153  *
 154  * The lowest driver-supported frequency which is equal or greater than the raw
 155  * next_freq (as calculated above) is returned, subject to policy min/max and
 156  * cpufreq driver limitations.
 157  */
 158 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
 159                                   unsigned long util, unsigned long max)
 160 {
 161         struct cpufreq_policy *policy = sg_policy->policy;
 162         unsigned int freq = arch_scale_freq_invariant() ?
 163                                 policy->cpuinfo.max_freq : policy->cur;
 164
 165         freq = map_util_freq(util, freq, max);
 166
 167         if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
 168                 return sg_policy->next_freq;
 169
 170         sg_policy->cached_raw_freq = freq;
 171         return cpufreq_driver_resolve_freq(policy, freq);
 172 }
 173
 174 /*
 175  * This function computes an effective utilization for the given CPU, to be
 176  * used for frequency selection given the linear relation: f = u * f_max.
 177  *
 178  * The scheduler tracks the following metrics:
 179  *
 180  *   cpu_util_{cfs,rt,dl,irq}()
 181  *   cpu_bw_dl()
 182  *
 183  * Where the cfs,rt and dl util numbers are tracked with the same metric and
 184  * synchronized windows and are thus directly comparable.
 185  *
 186  * The cfs,rt,dl utilization are the running times measured with rq->clock_task
 187  * which excludes things like IRQ and steal-time. These latter are then accrued
 188  * in the irq utilization.
 189  *
 190  * The DL bandwidth number otoh is not a measured metric but a value computed
 191  * based on the task model parameters and gives the minimal utilization
 192  * required to meet deadlines.
 193  */
 194 unsigned long schedutil_cpu_util(int cpu, unsigned long util_cfs,
 195                                  unsigned long max, enum schedutil_type type,
 196                                  struct task_struct *p)
 197 {
 198         unsigned long dl_util, util, irq;
 199         struct rq *rq = cpu_rq(cpu);
 200
 201         if (!uclamp_is_used() &&
 202             type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
 203                 return max;
 204         }
 205
 206         /*
 207          * Early check to see if IRQ/steal time saturates the CPU, can be
 208          * because of inaccuracies in how we track these -- see
 209          * update_irq_load_avg().
 210          */
 211         irq = cpu_util_irq(rq);
 212         if (unlikely(irq >= max))
 213                 return max;
 214
 215         /*
 216          * Because the time spend on RT/DL tasks is visible as 'lost' time to
 217          * CFS tasks and we use the same metric to track the effective
 218          * utilization (PELT windows are synchronized) we can directly add them
 219          * to obtain the CPU's actual utilization.
 220          *
 221          * CFS and RT utilization can be boosted or capped, depending on
 222          * utilization clamp constraints requested by currently RUNNABLE
 223          * tasks.
 224          * When there are no CFS RUNNABLE tasks, clamps are released and
 225          * frequency will be gracefully reduced with the utilization decay.
 226          */
 227         util = util_cfs + cpu_util_rt(rq);
 228         if (type == FREQUENCY_UTIL)
 229                 util = uclamp_rq_util_with(rq, util, p);
 230
 231         dl_util = cpu_util_dl(rq);
 232
 233         /*
 234          * For frequency selection we do not make cpu_util_dl() a permanent part
 235          * of this sum because we want to use cpu_bw_dl() later on, but we need
 236          * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
 237          * that we select f_max when there is no idle time.
 238          *
 239          * NOTE: numerical errors or stop class might cause us to not quite hit
 240          * saturation when we should -- something for later.
 241          */
 242         if (util + dl_util >= max)
 243                 return max;
 244
 245         /*
 246          * OTOH, for energy computation we need the estimated running time, so
 247          * include util_dl and ignore dl_bw.
 248          */
 249         if (type == ENERGY_UTIL)
 250                 util += dl_util;
 251
 252         /*
 253          * There is still idle time; further improve the number by using the
 254          * irq metric. Because IRQ/steal time is hidden from the task clock we
 255          * need to scale the task numbers:
 256          *
 257          *              max - irq
 258          *   U' = irq + --------- * U
 259          *                 max
 260          */
 261         util = scale_irq_capacity(util, irq, max);
 262         util += irq;
 263
 264         /*
 265          * Bandwidth required by DEADLINE must always be granted while, for
 266          * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
 267          * to gracefully reduce the frequency when no tasks show up for longer
 268          * periods of time.
 269          *
 270          * Ideally we would like to set bw_dl as min/guaranteed freq and util +
 271          * bw_dl as requested freq. However, cpufreq is not yet ready for such
 272          * an interface. So, we only do the latter for now.
 273          */
 274         if (type == FREQUENCY_UTIL)
 275                 util += cpu_bw_dl(rq);
 276
 277         return min(max, util);
 278 }
 279
 280 static void sugov_get_util(struct sugov_cpu *sg_cpu)
 281 {
 282         struct rq *rq = cpu_rq(sg_cpu->cpu);
 283         unsigned long max = arch_scale_cpu_capacity(sg_cpu->cpu);
 284
 285         sg_cpu->max = max;
 286         sg_cpu->bw_dl = cpu_bw_dl(rq);
 287         sg_cpu->util = schedutil_cpu_util(sg_cpu->cpu, cpu_util_cfs(rq), max,
 288                                           FREQUENCY_UTIL, NULL);
 289 }
 290
 291 /**
 292  * sugov_iowait_reset() - Reset the IO boost status of a CPU.
 293  * @sg_cpu: the sugov data for the CPU to boost
 294  * @time: the update time from the caller
 295  * @set_iowait_boost: true if an IO boost has been requested
 296  *
 297  * The IO wait boost of a task is disabled after a tick since the last update
 298  * of a CPU. If a new IO wait boost is requested after more then a tick, then
 299  * we enable the boost starting from IOWAIT_BOOST_MIN, which improves energy
 300  * efficiency by ignoring sporadic wakeups from IO.
 301  */
 302 static bool sugov_iowait_reset(struct sugov_cpu *sg_cpu, u64 time,
 303                                bool set_iowait_boost)
 304 {
 305         s64 delta_ns = time - sg_cpu->last_update;
 306
 307         /* Reset boost only if a tick has elapsed since last request */
 308         if (delta_ns <= TICK_NSEC)
 309                 return false;
 310
 311         sg_cpu->iowait_boost = set_iowait_boost ? IOWAIT_BOOST_MIN : 0;
 312         sg_cpu->iowait_boost_pending = set_iowait_boost;
 313
 314         return true;
 315 }
 316
 317 /**
 318  * sugov_iowait_boost() - Updates the IO boost status of a CPU.
 319  * @sg_cpu: the sugov data for the CPU to boost
 320  * @time: the update time from the caller
 321  * @flags: SCHED_CPUFREQ_IOWAIT if the task is waking up after an IO wait
 322  *
 323  * Each time a task wakes up after an IO operation, the CPU utilization can be
 324  * boosted to a certain utilization which doubles at each "frequent and
 325  * successive" wakeup from IO, ranging from IOWAIT_BOOST_MIN to the utilization
 326  * of the maximum OPP.
 327  *
 328  * To keep doubling, an IO boost has to be requested at least once per tick,
 329  * otherwise we restart from the utilization of the minimum OPP.
 330  */
 331 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
 332                                unsigned int flags)
 333 {
 334         bool set_iowait_boost = flags & SCHED_CPUFREQ_IOWAIT;
 335
 336         /* Reset boost if the CPU appears to have been idle enough */
 337         if (sg_cpu->iowait_boost &&
 338             sugov_iowait_reset(sg_cpu, time, set_iowait_boost))
 339                 return;
 340
 341         /* Boost only tasks waking up after IO */
 342         if (!set_iowait_boost)
 343                 return;
 344
 345         /* Ensure boost doubles only one time at each request */
 346         if (sg_cpu->iowait_boost_pending)
 347                 return;
 348         sg_cpu->iowait_boost_pending = true;
 349
 350         /* Double the boost at each request */
 351         if (sg_cpu->iowait_boost) {
 352                 sg_cpu->iowait_boost =
 353                         min_t(unsigned int, sg_cpu->iowait_boost << 1, SCHED_CAPACITY_SCALE);
 354                 return;
 355         }
 356
 357         /* First wakeup after IO: start with minimum boost */
 358         sg_cpu->iowait_boost = IOWAIT_BOOST_MIN;
 359 }
 360
 361 /**
 362  * sugov_iowait_apply() - Apply the IO boost to a CPU.
 363  * @sg_cpu: the sugov data for the cpu to boost
 364  * @time: the update time from the caller
 365  *
 366  * A CPU running a task which woken up after an IO operation can have its
 367  * utilization boosted to speed up the completion of those IO operations.
 368  * The IO boost value is increased each time a task wakes up from IO, in
 369  * sugov_iowait_apply(), and it's instead decreased by this function,
 370  * each time an increase has not been requested (!iowait_boost_pending).
 371  *
 372  * A CPU which also appears to have been idle for at least one tick has also
 373  * its IO boost utilization reset.
 374  *
 375  * This mechanism is designed to boost high frequently IO waiting tasks, while
 376  * being more conservative on tasks which does sporadic IO operations.
 377  */
 378 static void sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time)
 379 {
 380         unsigned long boost;
 381
 382         /* No boost currently required */
 383         if (!sg_cpu->iowait_boost)
 384                 return;
 385
 386         /* Reset boost if the CPU appears to have been idle enough */
 387         if (sugov_iowait_reset(sg_cpu, time, false))
 388                 return;
 389
 390         if (!sg_cpu->iowait_boost_pending) {
 391                 /*
 392                  * No boost pending; reduce the boost value.
 393                  */
 394                 sg_cpu->iowait_boost >>= 1;
 395                 if (sg_cpu->iowait_boost < IOWAIT_BOOST_MIN) {
 396                         sg_cpu->iowait_boost = 0;
 397                         return;
 398                 }
 399         }
 400
 401         sg_cpu->iowait_boost_pending = false;
 402
 403         /*
 404          * sg_cpu->util is already in capacity scale; convert iowait_boost
 405          * into the same scale so we can compare.
 406          */
 407         boost = (sg_cpu->iowait_boost * sg_cpu->max) >> SCHED_CAPACITY_SHIFT;
 408         if (sg_cpu->util < boost)
 409                 sg_cpu->util = boost;
 410 }
 411
 412 #ifdef CONFIG_NO_HZ_COMMON
 413 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
 414 {
 415         unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
 416         bool ret = idle_calls == sg_cpu->saved_idle_calls;
 417
 418         sg_cpu->saved_idle_calls = idle_calls;
 419         return ret;
 420 }
 421 #else
 422 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
 423 #endif /* CONFIG_NO_HZ_COMMON */
 424
 425 /*
 426  * Make sugov_should_update_freq() ignore the rate limit when DL
 427  * has increased the utilization.
 428  */
 429 static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu)
 430 {
 431         if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_dl)
 432                 sg_cpu->sg_policy->limits_changed = true;
 433 }
 434
 435 static inline bool sugov_update_single_common(struct sugov_cpu *sg_cpu,
 436                                               u64 time, unsigned int flags)
 437 {
 438         sugov_iowait_boost(sg_cpu, time, flags);
 439         sg_cpu->last_update = time;
 440
 441         ignore_dl_rate_limit(sg_cpu);
 442
 443         if (!sugov_should_update_freq(sg_cpu->sg_policy, time))
 444                 return false;
 445
 446         sugov_get_util(sg_cpu);
 447         sugov_iowait_apply(sg_cpu, time);
 448
 449         return true;
 450 }
 451
 452 static void sugov_update_single_freq(struct update_util_data *hook, u64 time,
 453                                      unsigned int flags)
 454 {
 455         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 456         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 457         unsigned int cached_freq = sg_policy->cached_raw_freq;
 458         unsigned int next_f;
 459
 460         if (!sugov_update_single_common(sg_cpu, time, flags))
 461                 return;
 462
 463         next_f = get_next_freq(sg_policy, sg_cpu->util, sg_cpu->max);
 464         /*
 465          * Do not reduce the frequency if the CPU has not been idle
 466          * recently, as the reduction is likely to be premature then.
 467          */
 468         if (sugov_cpu_is_busy(sg_cpu) && next_f < sg_policy->next_freq) {
 469                 next_f = sg_policy->next_freq;
 470
 471                 /* Restore cached freq as next_freq has changed */
 472                 sg_policy->cached_raw_freq = cached_freq;
 473         }
 474
 475         /*
 476          * This code runs under rq->lock for the target CPU, so it won't run
 477          * concurrently on two different CPUs for the same target and it is not
 478          * necessary to acquire the lock in the fast switch case.
 479          */
 480         if (sg_policy->policy->fast_switch_enabled) {
 481                 sugov_fast_switch(sg_policy, time, next_f);
 482         } else {
 483                 raw_spin_lock(&sg_policy->update_lock);
 484                 sugov_deferred_update(sg_policy, time, next_f);
 485                 raw_spin_unlock(&sg_policy->update_lock);
 486         }
 487 }
 488
 489 static void sugov_update_single_perf(struct update_util_data *hook, u64 time,
 490                                      unsigned int flags)
 491 {
 492         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 493         unsigned long prev_util = sg_cpu->util;
 494
 495         /*
 496          * Fall back to the "frequency" path if frequency invariance is not
 497          * supported, because the direct mapping between the utilization and
 498          * the performance levels depends on the frequency invariance.
 499          */
 500         if (!arch_scale_freq_invariant()) {
 501                 sugov_update_single_freq(hook, time, flags);
 502                 return;
 503         }
 504
 505         if (!sugov_update_single_common(sg_cpu, time, flags))
 506                 return;
 507
 508         /*
 509          * Do not reduce the target performance level if the CPU has not been
 510          * idle recently, as the reduction is likely to be premature then.
 511          */
 512         if (sugov_cpu_is_busy(sg_cpu) && sg_cpu->util < prev_util)
 513                 sg_cpu->util = prev_util;
 514
 515         cpufreq_driver_adjust_perf(sg_cpu->cpu, map_util_perf(sg_cpu->bw_dl),
 516                                    map_util_perf(sg_cpu->util), sg_cpu->max);
 517
 518         sg_cpu->sg_policy->last_freq_update_time = time;
 519 }
 520
 521 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
 522 {
 523         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 524         struct cpufreq_policy *policy = sg_policy->policy;
 525         unsigned long util = 0, max = 1;
 526         unsigned int j;
 527
 528         for_each_cpu(j, policy->cpus) {
 529                 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
 530                 unsigned long j_util, j_max;
 531
 532                 sugov_get_util(j_sg_cpu);
 533                 sugov_iowait_apply(j_sg_cpu, time);
 534                 j_util = j_sg_cpu->util;
 535                 j_max = j_sg_cpu->max;
 536
 537                 if (j_util * max > j_max * util) {
 538                         util = j_util;
 539                         max = j_max;
 540                 }
 541         }
 542
 543         return get_next_freq(sg_policy, util, max);
 544 }
 545
 546 static void
 547 sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
 548 {
 549         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 550         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 551         unsigned int next_f;
 552
 553         raw_spin_lock(&sg_policy->update_lock);
 554
 555         sugov_iowait_boost(sg_cpu, time, flags);
 556         sg_cpu->last_update = time;
 557
 558         ignore_dl_rate_limit(sg_cpu);
 559
 560         if (sugov_should_update_freq(sg_policy, time)) {
 561                 next_f = sugov_next_freq_shared(sg_cpu, time);
 562
 563                 if (sg_policy->policy->fast_switch_enabled)
 564                         sugov_fast_switch(sg_policy, time, next_f);
 565                 else
 566                         sugov_deferred_update(sg_policy, time, next_f);
 567         }
 568
 569         raw_spin_unlock(&sg_policy->update_lock);
 570 }
 571
 572 static void sugov_work(struct kthread_work *work)
 573 {
 574         struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
 575         unsigned int freq;
 576         unsigned long flags;
 577
 578         /*
 579          * Hold sg_policy->update_lock shortly to handle the case where:
 580          * incase sg_policy->next_freq is read here, and then updated by
 581          * sugov_deferred_update() just before work_in_progress is set to false
 582          * here, we may miss queueing the new update.
 583          *
 584          * Note: If a work was queued after the update_lock is released,
 585          * sugov_work() will just be called again by kthread_work code; and the
 586          * request will be proceed before the sugov thread sleeps.
 587          */
 588         raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
 589         freq = sg_policy->next_freq;
 590         sg_policy->work_in_progress = false;
 591         raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
 592
 593         mutex_lock(&sg_policy->work_lock);
 594         __cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
 595         mutex_unlock(&sg_policy->work_lock);
 596 }
 597
 598 static void sugov_irq_work(struct irq_work *irq_work)
 599 {
 600         struct sugov_policy *sg_policy;
 601
 602         sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
 603
 604         kthread_queue_work(&sg_policy->worker, &sg_policy->work);
 605 }
 606
 607 /************************** sysfs interface ************************/
 608
 609 static struct sugov_tunables *global_tunables;
 610 static DEFINE_MUTEX(global_tunables_lock);
 611
 612 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
 613 {
 614         return container_of(attr_set, struct sugov_tunables, attr_set);
 615 }
 616
 617 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
 618 {
 619         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 620
 621         return sprintf(buf, "%u\n", tunables->rate_limit_us);
 622 }
 623
 624 static ssize_t
 625 rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
 626 {
 627         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 628         struct sugov_policy *sg_policy;
 629         unsigned int rate_limit_us;
 630
 631         if (kstrtouint(buf, 10, &rate_limit_us))
 632                 return -EINVAL;
 633
 634         tunables->rate_limit_us = rate_limit_us;
 635
 636         list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
 637                 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
 638
 639         return count;
 640 }
 641
 642 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
 643
 644 static struct attribute *sugov_attrs[] = {
 645         &rate_limit_us.attr,
 646         NULL
 647 };
 648 ATTRIBUTE_GROUPS(sugov);
 649
 650 static struct kobj_type sugov_tunables_ktype = {
 651         .default_groups = sugov_groups,
 652         .sysfs_ops = &governor_sysfs_ops,
 653 };
 654
 655 /********************** cpufreq governor interface *********************/
 656
 657 struct cpufreq_governor schedutil_gov;
 658
 659 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
 660 {
 661         struct sugov_policy *sg_policy;
 662
 663         sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
 664         if (!sg_policy)
 665                 return NULL;
 666
 667         sg_policy->policy = policy;
 668         raw_spin_lock_init(&sg_policy->update_lock);
 669         return sg_policy;
 670 }
 671
 672 static void sugov_policy_free(struct sugov_policy *sg_policy)
 673 {
 674         kfree(sg_policy);
 675 }
 676
 677 static int sugov_kthread_create(struct sugov_policy *sg_policy)
 678 {
 679         struct task_struct *thread;
 680         struct sched_attr attr = {
 681                 .size           = sizeof(struct sched_attr),
 682                 .sched_policy   = SCHED_DEADLINE,
 683                 .sched_flags    = SCHED_FLAG_SUGOV,
 684                 .sched_nice     = 0,
 685                 .sched_priority = 0,
 686                 /*
 687                  * Fake (unused) bandwidth; workaround to "fix"
 688                  * priority inheritance.
 689                  */
 690                 .sched_runtime  =  1000000,
 691                 .sched_deadline = 10000000,
 692                 .sched_period   = 10000000,
 693         };
 694         struct cpufreq_policy *policy = sg_policy->policy;
 695         int ret;
 696
 697         /* kthread only required for slow path */
 698         if (policy->fast_switch_enabled)
 699                 return 0;
 700
 701         kthread_init_work(&sg_policy->work, sugov_work);
 702         kthread_init_worker(&sg_policy->worker);
 703         thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
 704                                 "sugov:%d",
 705                                 cpumask_first(policy->related_cpus));
 706         if (IS_ERR(thread)) {
 707                 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
 708                 return PTR_ERR(thread);
 709         }
 710
 711         ret = sched_setattr_nocheck(thread, &attr);
 712         if (ret) {
 713                 kthread_stop(thread);
 714                 pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
 715                 return ret;
 716         }
 717
 718         sg_policy->thread = thread;
 719         kthread_bind_mask(thread, policy->related_cpus);
 720         init_irq_work(&sg_policy->irq_work, sugov_irq_work);
 721         mutex_init(&sg_policy->work_lock);
 722
 723         wake_up_process(thread);
 724
 725         return 0;
 726 }
 727
 728 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
 729 {
 730         /* kthread only required for slow path */
 731         if (sg_policy->policy->fast_switch_enabled)
 732                 return;
 733
 734         kthread_flush_worker(&sg_policy->worker);
 735         kthread_stop(sg_policy->thread);
 736         mutex_destroy(&sg_policy->work_lock);
 737 }
 738
 739 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
 740 {
 741         struct sugov_tunables *tunables;
 742
 743         tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
 744         if (tunables) {
 745                 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
 746                 if (!have_governor_per_policy())
 747                         global_tunables = tunables;
 748         }
 749         return tunables;
 750 }
 751
 752 static void sugov_tunables_free(struct sugov_tunables *tunables)
 753 {
 754         if (!have_governor_per_policy())
 755                 global_tunables = NULL;
 756
 757         kfree(tunables);
 758 }
 759
 760 static int sugov_init(struct cpufreq_policy *policy)
 761 {
 762         struct sugov_policy *sg_policy;
 763         struct sugov_tunables *tunables;
 764         int ret = 0;
 765
 766         /* State should be equivalent to EXIT */
 767         if (policy->governor_data)
 768                 return -EBUSY;
 769
 770         cpufreq_enable_fast_switch(policy);
 771
 772         sg_policy = sugov_policy_alloc(policy);
 773         if (!sg_policy) {
 774                 ret = -ENOMEM;
 775                 goto disable_fast_switch;
 776         }
 777
 778         ret = sugov_kthread_create(sg_policy);
 779         if (ret)
 780                 goto free_sg_policy;
 781
 782         mutex_lock(&global_tunables_lock);
 783
 784         if (global_tunables) {
 785                 if (WARN_ON(have_governor_per_policy())) {
 786                         ret = -EINVAL;
 787                         goto stop_kthread;
 788                 }
 789                 policy->governor_data = sg_policy;
 790                 sg_policy->tunables = global_tunables;
 791
 792                 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
 793                 goto out;
 794         }
 795
 796         tunables = sugov_tunables_alloc(sg_policy);
 797         if (!tunables) {
 798                 ret = -ENOMEM;
 799                 goto stop_kthread;
 800         }
 801
 802         tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
 803
 804         policy->governor_data = sg_policy;
 805         sg_policy->tunables = tunables;
 806
 807         ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
 808                                    get_governor_parent_kobj(policy), "%s",
 809                                    schedutil_gov.name);
 810         if (ret)
 811                 goto fail;
 812
 813 out:
 814         mutex_unlock(&global_tunables_lock);
 815         return 0;
 816
 817 fail:
 818         kobject_put(&tunables->attr_set.kobj);
 819         policy->governor_data = NULL;
 820         sugov_tunables_free(tunables);
 821
 822 stop_kthread:
 823         sugov_kthread_stop(sg_policy);
 824         mutex_unlock(&global_tunables_lock);
 825
 826 free_sg_policy:
 827         sugov_policy_free(sg_policy);
 828
 829 disable_fast_switch:
 830         cpufreq_disable_fast_switch(policy);
 831
 832         pr_err("initialization failed (error %d)\n", ret);
 833         return ret;
 834 }
 835
 836 static void sugov_exit(struct cpufreq_policy *policy)
 837 {
 838         struct sugov_policy *sg_policy = policy->governor_data;
 839         struct sugov_tunables *tunables = sg_policy->tunables;
 840         unsigned int count;
 841
 842         mutex_lock(&global_tunables_lock);
 843
 844         count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
 845         policy->governor_data = NULL;
 846         if (!count)
 847                 sugov_tunables_free(tunables);
 848
 849         mutex_unlock(&global_tunables_lock);
 850
 851         sugov_kthread_stop(sg_policy);
 852         sugov_policy_free(sg_policy);
 853         cpufreq_disable_fast_switch(policy);
 854 }
 855
 856 static int sugov_start(struct cpufreq_policy *policy)
 857 {
 858         struct sugov_policy *sg_policy = policy->governor_data;
 859         void (*uu)(struct update_util_data *data, u64 time, unsigned int flags);
 860         unsigned int cpu;
 861
 862         sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
 863         sg_policy->last_freq_update_time        = 0;
 864         sg_policy->next_freq                    = 0;
 865         sg_policy->work_in_progress             = false;
 866         sg_policy->limits_changed               = false;
 867         sg_policy->cached_raw_freq              = 0;
 868
 869         sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
 870
 871         for_each_cpu(cpu, policy->cpus) {
 872                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 873
 874                 memset(sg_cpu, 0, sizeof(*sg_cpu));
 875                 sg_cpu->cpu                     = cpu;
 876                 sg_cpu->sg_policy               = sg_policy;
 877         }
 878
 879         if (policy_is_shared(policy))
 880                 uu = sugov_update_shared;
 881         else if (policy->fast_switch_enabled && cpufreq_driver_has_adjust_perf())
 882                 uu = sugov_update_single_perf;
 883         else
 884                 uu = sugov_update_single_freq;
 885
 886         for_each_cpu(cpu, policy->cpus) {
 887                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 888
 889                 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util, uu);
 890         }
 891         return 0;
 892 }
 893
 894 static void sugov_stop(struct cpufreq_policy *policy)
 895 {
 896         struct sugov_policy *sg_policy = policy->governor_data;
 897         unsigned int cpu;
 898
 899         for_each_cpu(cpu, policy->cpus)
 900                 cpufreq_remove_update_util_hook(cpu);
 901
 902         synchronize_rcu();
 903
 904         if (!policy->fast_switch_enabled) {
 905                 irq_work_sync(&sg_policy->irq_work);
 906                 kthread_cancel_work_sync(&sg_policy->work);
 907         }
 908 }
 909
 910 static void sugov_limits(struct cpufreq_policy *policy)
 911 {
 912         struct sugov_policy *sg_policy = policy->governor_data;
 913
 914         if (!policy->fast_switch_enabled) {
 915                 mutex_lock(&sg_policy->work_lock);
 916                 cpufreq_policy_apply_limits(policy);
 917                 mutex_unlock(&sg_policy->work_lock);
 918         }
 919
 920         sg_policy->limits_changed = true;
 921 }
 922
 923 struct cpufreq_governor schedutil_gov = {
 924         .name                   = "schedutil",
 925         .owner                  = THIS_MODULE,
 926         .flags                  = CPUFREQ_GOV_DYNAMIC_SWITCHING,
 927         .init                   = sugov_init,
 928         .exit                   = sugov_exit,
 929         .start                  = sugov_start,
 930         .stop                   = sugov_stop,
 931         .limits                 = sugov_limits,
 932 };
 933
 934 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
 935 struct cpufreq_governor *cpufreq_default_governor(void)
 936 {
 937         return &schedutil_gov;
 938 }
 939 #endif
 940
 941 cpufreq_governor_init(schedutil_gov);
 942
 943 #ifdef CONFIG_ENERGY_MODEL
 944 static void rebuild_sd_workfn(struct work_struct *work)
 945 {
 946         rebuild_sched_domains_energy();
 947 }
 948 static DECLARE_WORK(rebuild_sd_work, rebuild_sd_workfn);
 949
 950 /*
 951  * EAS shouldn't be attempted without sugov, so rebuild the sched_domains
 952  * on governor changes to make sure the scheduler knows about it.
 953  */
 954 void sched_cpufreq_governor_change(struct cpufreq_policy *policy,
 955                                   struct cpufreq_governor *old_gov)
 956 {
 957         if (old_gov == &schedutil_gov || policy->governor == &schedutil_gov) {
 958                 /*
 959                  * When called from the cpufreq_register_driver() path, the
 960                  * cpu_hotplug_lock is already held, so use a work item to
 961                  * avoid nested locking in rebuild_sched_domains().
 962                  */
 963                 schedule_work(&rebuild_sd_work);
 964         }
 965
 966 }
 967 #endif