kernel/sched/cpufreq_schedutil.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * CPUFreq governor based on scheduler-provided CPU utilization data.
   4  *
   5  * Copyright (C) 2016, Intel Corporation
   6  * Author: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
   7  */
   8
   9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  11 #include "sched.h"
  12
  13 #include <linux/sched/cpufreq.h>
  14 #include <trace/events/power.h>
  15
  16 #define IOWAIT_BOOST_MIN        (SCHED_CAPACITY_SCALE / 8)
  17
  18 struct sugov_tunables {
  19         struct gov_attr_set     attr_set;
  20         unsigned int            rate_limit_us;
  21 };
  22
  23 struct sugov_policy {
  24         struct cpufreq_policy   *policy;
  25
  26         struct sugov_tunables   *tunables;
  27         struct list_head        tunables_hook;
  28
  29         raw_spinlock_t          update_lock;    /* For shared policies */
  30         u64                     last_freq_update_time;
  31         s64                     freq_update_delay_ns;
  32         unsigned int            next_freq;
  33         unsigned int            cached_raw_freq;
  34
  35         /* The next fields are only needed if fast switch cannot be used: */
  36         struct                  irq_work irq_work;
  37         struct                  kthread_work work;
  38         struct                  mutex work_lock;
  39         struct                  kthread_worker worker;
  40         struct task_struct      *thread;
  41         bool                    work_in_progress;
  42
  43         bool                    limits_changed;
  44         bool                    need_freq_update;
  45 };
  46
  47 struct sugov_cpu {
  48         struct update_util_data update_util;
  49         struct sugov_policy     *sg_policy;
  50         unsigned int            cpu;
  51
  52         bool                    iowait_boost_pending;
  53         unsigned int            iowait_boost;
  54         u64                     last_update;
  55
  56         unsigned long           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, struct sugov_policy *sg_policy)
 430 {
 431         if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_dl)
 432                 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         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 439
 440         sugov_iowait_boost(sg_cpu, time, flags);
 441         sg_cpu->last_update = time;
 442
 443         ignore_dl_rate_limit(sg_cpu, sg_policy);
 444
 445         if (!sugov_should_update_freq(sg_policy, time))
 446                 return false;
 447
 448         sugov_get_util(sg_cpu);
 449         sugov_iowait_apply(sg_cpu, time);
 450
 451         return true;
 452 }
 453
 454 static void sugov_update_single_freq(struct update_util_data *hook, u64 time,
 455                                      unsigned int flags)
 456 {
 457         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 458         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 459         unsigned int cached_freq = sg_policy->cached_raw_freq;
 460         unsigned int next_f;
 461
 462         if (!sugov_update_single_common(sg_cpu, time, flags))
 463                 return;
 464
 465         next_f = get_next_freq(sg_policy, sg_cpu->util, sg_cpu->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 (sugov_cpu_is_busy(sg_cpu) && next_f < sg_policy->next_freq) {
 471                 next_f = sg_policy->next_freq;
 472
 473                 /* Restore cached freq as next_freq has changed */
 474                 sg_policy->cached_raw_freq = cached_freq;
 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 void sugov_update_single_perf(struct update_util_data *hook, u64 time,
 492                                      unsigned int flags)
 493 {
 494         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 495         unsigned long prev_util = sg_cpu->util;
 496
 497         /*
 498          * Fall back to the "frequency" path if frequency invariance is not
 499          * supported, because the direct mapping between the utilization and
 500          * the performance levels depends on the frequency invariance.
 501          */
 502         if (!arch_scale_freq_invariant()) {
 503                 sugov_update_single_freq(hook, time, flags);
 504                 return;
 505         }
 506
 507         if (!sugov_update_single_common(sg_cpu, time, flags))
 508                 return;
 509
 510         /*
 511          * Do not reduce the target performance level if the CPU has not been
 512          * idle recently, as the reduction is likely to be premature then.
 513          */
 514         if (sugov_cpu_is_busy(sg_cpu) && sg_cpu->util < prev_util)
 515                 sg_cpu->util = prev_util;
 516
 517         cpufreq_driver_adjust_perf(sg_cpu->cpu, map_util_perf(sg_cpu->bw_dl),
 518                                    map_util_perf(sg_cpu->util), sg_cpu->max);
 519
 520         sg_cpu->sg_policy->last_freq_update_time = time;
 521 }
 522
 523 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
 524 {
 525         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 526         struct cpufreq_policy *policy = sg_policy->policy;
 527         unsigned long util = 0, max = 1;
 528         unsigned int j;
 529
 530         for_each_cpu(j, policy->cpus) {
 531                 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
 532                 unsigned long j_util, j_max;
 533
 534                 sugov_get_util(j_sg_cpu);
 535                 sugov_iowait_apply(j_sg_cpu, time);
 536                 j_util = j_sg_cpu->util;
 537                 j_max = j_sg_cpu->max;
 538
 539                 if (j_util * max > j_max * util) {
 540                         util = j_util;
 541                         max = j_max;
 542                 }
 543         }
 544
 545         return get_next_freq(sg_policy, util, max);
 546 }
 547
 548 static void
 549 sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
 550 {
 551         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 552         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 553         unsigned int next_f;
 554
 555         raw_spin_lock(&sg_policy->update_lock);
 556
 557         sugov_iowait_boost(sg_cpu, time, flags);
 558         sg_cpu->last_update = time;
 559
 560         ignore_dl_rate_limit(sg_cpu, sg_policy);
 561
 562         if (sugov_should_update_freq(sg_policy, time)) {
 563                 next_f = sugov_next_freq_shared(sg_cpu, time);
 564
 565                 if (sg_policy->policy->fast_switch_enabled)
 566                         sugov_fast_switch(sg_policy, time, next_f);
 567                 else
 568                         sugov_deferred_update(sg_policy, time, next_f);
 569         }
 570
 571         raw_spin_unlock(&sg_policy->update_lock);
 572 }
 573
 574 static void sugov_work(struct kthread_work *work)
 575 {
 576         struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
 577         unsigned int freq;
 578         unsigned long flags;
 579
 580         /*
 581          * Hold sg_policy->update_lock shortly to handle the case where:
 582          * incase sg_policy->next_freq is read here, and then updated by
 583          * sugov_deferred_update() just before work_in_progress is set to false
 584          * here, we may miss queueing the new update.
 585          *
 586          * Note: If a work was queued after the update_lock is released,
 587          * sugov_work() will just be called again by kthread_work code; and the
 588          * request will be proceed before the sugov thread sleeps.
 589          */
 590         raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
 591         freq = sg_policy->next_freq;
 592         sg_policy->work_in_progress = false;
 593         raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
 594
 595         mutex_lock(&sg_policy->work_lock);
 596         __cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
 597         mutex_unlock(&sg_policy->work_lock);
 598 }
 599
 600 static void sugov_irq_work(struct irq_work *irq_work)
 601 {
 602         struct sugov_policy *sg_policy;
 603
 604         sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
 605
 606         kthread_queue_work(&sg_policy->worker, &sg_policy->work);
 607 }
 608
 609 /************************** sysfs interface ************************/
 610
 611 static struct sugov_tunables *global_tunables;
 612 static DEFINE_MUTEX(global_tunables_lock);
 613
 614 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
 615 {
 616         return container_of(attr_set, struct sugov_tunables, attr_set);
 617 }
 618
 619 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
 620 {
 621         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 622
 623         return sprintf(buf, "%u\n", tunables->rate_limit_us);
 624 }
 625
 626 static ssize_t
 627 rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
 628 {
 629         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 630         struct sugov_policy *sg_policy;
 631         unsigned int rate_limit_us;
 632
 633         if (kstrtouint(buf, 10, &rate_limit_us))
 634                 return -EINVAL;
 635
 636         tunables->rate_limit_us = rate_limit_us;
 637
 638         list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
 639                 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
 640
 641         return count;
 642 }
 643
 644 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
 645
 646 static struct attribute *sugov_attrs[] = {
 647         &rate_limit_us.attr,
 648         NULL
 649 };
 650 ATTRIBUTE_GROUPS(sugov);
 651
 652 static struct kobj_type sugov_tunables_ktype = {
 653         .default_groups = sugov_groups,
 654         .sysfs_ops = &governor_sysfs_ops,
 655 };
 656
 657 /********************** cpufreq governor interface *********************/
 658
 659 struct cpufreq_governor schedutil_gov;
 660
 661 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
 662 {
 663         struct sugov_policy *sg_policy;
 664
 665         sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
 666         if (!sg_policy)
 667                 return NULL;
 668
 669         sg_policy->policy = policy;
 670         raw_spin_lock_init(&sg_policy->update_lock);
 671         return sg_policy;
 672 }
 673
 674 static void sugov_policy_free(struct sugov_policy *sg_policy)
 675 {
 676         kfree(sg_policy);
 677 }
 678
 679 static int sugov_kthread_create(struct sugov_policy *sg_policy)
 680 {
 681         struct task_struct *thread;
 682         struct sched_attr attr = {
 683                 .size           = sizeof(struct sched_attr),
 684                 .sched_policy   = SCHED_DEADLINE,
 685                 .sched_flags    = SCHED_FLAG_SUGOV,
 686                 .sched_nice     = 0,
 687                 .sched_priority = 0,
 688                 /*
 689                  * Fake (unused) bandwidth; workaround to "fix"
 690                  * priority inheritance.
 691                  */
 692                 .sched_runtime  =  1000000,
 693                 .sched_deadline = 10000000,
 694                 .sched_period   = 10000000,
 695         };
 696         struct cpufreq_policy *policy = sg_policy->policy;
 697         int ret;
 698
 699         /* kthread only required for slow path */
 700         if (policy->fast_switch_enabled)
 701                 return 0;
 702
 703         kthread_init_work(&sg_policy->work, sugov_work);
 704         kthread_init_worker(&sg_policy->worker);
 705         thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
 706                                 "sugov:%d",
 707                                 cpumask_first(policy->related_cpus));
 708         if (IS_ERR(thread)) {
 709                 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
 710                 return PTR_ERR(thread);
 711         }
 712
 713         ret = sched_setattr_nocheck(thread, &attr);
 714         if (ret) {
 715                 kthread_stop(thread);
 716                 pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
 717                 return ret;
 718         }
 719
 720         sg_policy->thread = thread;
 721         kthread_bind_mask(thread, policy->related_cpus);
 722         init_irq_work(&sg_policy->irq_work, sugov_irq_work);
 723         mutex_init(&sg_policy->work_lock);
 724
 725         wake_up_process(thread);
 726
 727         return 0;
 728 }
 729
 730 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
 731 {
 732         /* kthread only required for slow path */
 733         if (sg_policy->policy->fast_switch_enabled)
 734                 return;
 735
 736         kthread_flush_worker(&sg_policy->worker);
 737         kthread_stop(sg_policy->thread);
 738         mutex_destroy(&sg_policy->work_lock);
 739 }
 740
 741 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
 742 {
 743         struct sugov_tunables *tunables;
 744
 745         tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
 746         if (tunables) {
 747                 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
 748                 if (!have_governor_per_policy())
 749                         global_tunables = tunables;
 750         }
 751         return tunables;
 752 }
 753
 754 static void sugov_tunables_free(struct sugov_tunables *tunables)
 755 {
 756         if (!have_governor_per_policy())
 757                 global_tunables = NULL;
 758
 759         kfree(tunables);
 760 }
 761
 762 static int sugov_init(struct cpufreq_policy *policy)
 763 {
 764         struct sugov_policy *sg_policy;
 765         struct sugov_tunables *tunables;
 766         int ret = 0;
 767
 768         /* State should be equivalent to EXIT */
 769         if (policy->governor_data)
 770                 return -EBUSY;
 771
 772         cpufreq_enable_fast_switch(policy);
 773
 774         sg_policy = sugov_policy_alloc(policy);
 775         if (!sg_policy) {
 776                 ret = -ENOMEM;
 777                 goto disable_fast_switch;
 778         }
 779
 780         ret = sugov_kthread_create(sg_policy);
 781         if (ret)
 782                 goto free_sg_policy;
 783
 784         mutex_lock(&global_tunables_lock);
 785
 786         if (global_tunables) {
 787                 if (WARN_ON(have_governor_per_policy())) {
 788                         ret = -EINVAL;
 789                         goto stop_kthread;
 790                 }
 791                 policy->governor_data = sg_policy;
 792                 sg_policy->tunables = global_tunables;
 793
 794                 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
 795                 goto out;
 796         }
 797
 798         tunables = sugov_tunables_alloc(sg_policy);
 799         if (!tunables) {
 800                 ret = -ENOMEM;
 801                 goto stop_kthread;
 802         }
 803
 804         tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
 805
 806         policy->governor_data = sg_policy;
 807         sg_policy->tunables = tunables;
 808
 809         ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
 810                                    get_governor_parent_kobj(policy), "%s",
 811                                    schedutil_gov.name);
 812         if (ret)
 813                 goto fail;
 814
 815 out:
 816         mutex_unlock(&global_tunables_lock);
 817         return 0;
 818
 819 fail:
 820         kobject_put(&tunables->attr_set.kobj);
 821         policy->governor_data = NULL;
 822         sugov_tunables_free(tunables);
 823
 824 stop_kthread:
 825         sugov_kthread_stop(sg_policy);
 826         mutex_unlock(&global_tunables_lock);
 827
 828 free_sg_policy:
 829         sugov_policy_free(sg_policy);
 830
 831 disable_fast_switch:
 832         cpufreq_disable_fast_switch(policy);
 833
 834         pr_err("initialization failed (error %d)\n", ret);
 835         return ret;
 836 }
 837
 838 static void sugov_exit(struct cpufreq_policy *policy)
 839 {
 840         struct sugov_policy *sg_policy = policy->governor_data;
 841         struct sugov_tunables *tunables = sg_policy->tunables;
 842         unsigned int count;
 843
 844         mutex_lock(&global_tunables_lock);
 845
 846         count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
 847         policy->governor_data = NULL;
 848         if (!count)
 849                 sugov_tunables_free(tunables);
 850
 851         mutex_unlock(&global_tunables_lock);
 852
 853         sugov_kthread_stop(sg_policy);
 854         sugov_policy_free(sg_policy);
 855         cpufreq_disable_fast_switch(policy);
 856 }
 857
 858 static int sugov_start(struct cpufreq_policy *policy)
 859 {
 860         struct sugov_policy *sg_policy = policy->governor_data;
 861         void (*uu)(struct update_util_data *data, u64 time, unsigned int flags);
 862         unsigned int cpu;
 863
 864         sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
 865         sg_policy->last_freq_update_time        = 0;
 866         sg_policy->next_freq                    = 0;
 867         sg_policy->work_in_progress             = false;
 868         sg_policy->limits_changed               = false;
 869         sg_policy->cached_raw_freq              = 0;
 870
 871         sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
 872
 873         for_each_cpu(cpu, policy->cpus) {
 874                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 875
 876                 memset(sg_cpu, 0, sizeof(*sg_cpu));
 877                 sg_cpu->cpu                     = cpu;
 878                 sg_cpu->sg_policy               = sg_policy;
 879         }
 880
 881         if (policy_is_shared(policy))
 882                 uu = sugov_update_shared;
 883         else if (policy->fast_switch_enabled && cpufreq_driver_has_adjust_perf())
 884                 uu = sugov_update_single_perf;
 885         else
 886                 uu = sugov_update_single_freq;
 887
 888         for_each_cpu(cpu, policy->cpus) {
 889                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 890
 891                 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util, uu);
 892         }
 893         return 0;
 894 }
 895
 896 static void sugov_stop(struct cpufreq_policy *policy)
 897 {
 898         struct sugov_policy *sg_policy = policy->governor_data;
 899         unsigned int cpu;
 900
 901         for_each_cpu(cpu, policy->cpus)
 902                 cpufreq_remove_update_util_hook(cpu);
 903
 904         synchronize_rcu();
 905
 906         if (!policy->fast_switch_enabled) {
 907                 irq_work_sync(&sg_policy->irq_work);
 908                 kthread_cancel_work_sync(&sg_policy->work);
 909         }
 910 }
 911
 912 static void sugov_limits(struct cpufreq_policy *policy)
 913 {
 914         struct sugov_policy *sg_policy = policy->governor_data;
 915
 916         if (!policy->fast_switch_enabled) {
 917                 mutex_lock(&sg_policy->work_lock);
 918                 cpufreq_policy_apply_limits(policy);
 919                 mutex_unlock(&sg_policy->work_lock);
 920         }
 921
 922         sg_policy->limits_changed = true;
 923 }
 924
 925 struct cpufreq_governor schedutil_gov = {
 926         .name                   = "schedutil",
 927         .owner                  = THIS_MODULE,
 928         .flags                  = CPUFREQ_GOV_DYNAMIC_SWITCHING,
 929         .init                   = sugov_init,
 930         .exit                   = sugov_exit,
 931         .start                  = sugov_start,
 932         .stop                   = sugov_stop,
 933         .limits                 = sugov_limits,
 934 };
 935
 936 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
 937 struct cpufreq_governor *cpufreq_default_governor(void)
 938 {
 939         return &schedutil_gov;
 940 }
 941 #endif
 942
 943 cpufreq_governor_init(schedutil_gov);
 944
 945 #ifdef CONFIG_ENERGY_MODEL
 946 static void rebuild_sd_workfn(struct work_struct *work)
 947 {
 948         rebuild_sched_domains_energy();
 949 }
 950 static DECLARE_WORK(rebuild_sd_work, rebuild_sd_workfn);
 951
 952 /*
 953  * EAS shouldn't be attempted without sugov, so rebuild the sched_domains
 954  * on governor changes to make sure the scheduler knows about it.
 955  */
 956 void sched_cpufreq_governor_change(struct cpufreq_policy *policy,
 957                                   struct cpufreq_governor *old_gov)
 958 {
 959         if (old_gov == &schedutil_gov || policy->governor == &schedutil_gov) {
 960                 /*
 961                  * When called from the cpufreq_register_driver() path, the
 962                  * cpu_hotplug_lock is already held, so use a work item to
 963                  * avoid nested locking in rebuild_sched_domains().
 964                  */
 965                 schedule_work(&rebuild_sd_work);
 966         }
 967
 968 }
 969 #endif