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 <[email protected]>
   7  */
   8
   9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  11 #include "sched.h"
  12
  13 #include <linux/sched/cpufreq.h>
  14 #include <trace/events/power.h>
  15
  16 #define IOWAIT_BOOST_MIN        (SCHED_CAPACITY_SCALE / 8)
  17
  18 struct sugov_tunables {
  19         struct gov_attr_set     attr_set;
  20         unsigned int            rate_limit_us;
  21 };
  22
  23 struct sugov_policy {
  24         struct cpufreq_policy   *policy;
  25
  26         struct sugov_tunables   *tunables;
  27         struct list_head        tunables_hook;
  28
  29         raw_spinlock_t          update_lock;    /* For shared policies */
  30         u64                     last_freq_update_time;
  31         s64                     freq_update_delay_ns;
  32         unsigned int            next_freq;
  33         unsigned int            cached_raw_freq;
  34
  35         /* The next fields are only needed if fast switch cannot be used: */
  36         struct                  irq_work irq_work;
  37         struct                  kthread_work work;
  38         struct                  mutex work_lock;
  39         struct                  kthread_worker worker;
  40         struct task_struct      *thread;
  41         bool                    work_in_progress;
  42
  43         bool                    limits_changed;
  44         bool                    need_freq_update;
  45 };
  46
  47 struct sugov_cpu {
  48         struct update_util_data update_util;
  49         struct sugov_policy     *sg_policy;
  50         unsigned int            cpu;
  51
  52         bool                    iowait_boost_pending;
  53         unsigned int            iowait_boost;
  54         u64                     last_update;
  55
  56         unsigned long           bw_dl;
  57         unsigned long           max;
  58
  59         /* The field below is for single-CPU policies only: */
  60 #ifdef CONFIG_NO_HZ_COMMON
  61         unsigned long           saved_idle_calls;
  62 #endif
  63 };
  64
  65 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
  66
  67 /************************ Governor internals ***********************/
  68
  69 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
  70 {
  71         s64 delta_ns;
  72
  73         /*
  74          * Since cpufreq_update_util() is called with rq->lock held for
  75          * the @target_cpu, our per-CPU data is fully serialized.
  76          *
  77          * However, drivers cannot in general deal with cross-CPU
  78          * requests, so while get_next_freq() will work, our
  79          * sugov_update_commit() call may not for the fast switching platforms.
  80          *
  81          * Hence stop here for remote requests if they aren't supported
  82          * by the hardware, as calculating the frequency is pointless if
  83          * we cannot in fact act on it.
  84          *
  85          * This is needed on the slow switching platforms too to prevent CPUs
  86          * going offline from leaving stale IRQ work items behind.
  87          */
  88         if (!cpufreq_this_cpu_can_update(sg_policy->policy))
  89                 return false;
  90
  91         if (unlikely(sg_policy->limits_changed)) {
  92                 sg_policy->limits_changed = false;
  93                 sg_policy->need_freq_update = true;
  94                 return true;
  95         }
  96
  97         delta_ns = time - sg_policy->last_freq_update_time;
  98
  99         return delta_ns >= sg_policy->freq_update_delay_ns;
 100 }
 101
 102 static bool sugov_update_next_freq(struct sugov_policy *sg_policy, u64 time,
 103                                    unsigned int next_freq)
 104 {
 105         if (sg_policy->need_freq_update)
 106                 sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
 107         else if (sg_policy->next_freq == next_freq)
 108                 return false;
 109
 110         sg_policy->next_freq = next_freq;
 111         sg_policy->last_freq_update_time = time;
 112
 113         return true;
 114 }
 115
 116 static void sugov_fast_switch(struct sugov_policy *sg_policy, u64 time,
 117                               unsigned int next_freq)
 118 {
 119         if (sugov_update_next_freq(sg_policy, time, next_freq))
 120                 cpufreq_driver_fast_switch(sg_policy->policy, next_freq);
 121 }
 122
 123 static void sugov_deferred_update(struct sugov_policy *sg_policy, u64 time,
 124                                   unsigned int next_freq)
 125 {
 126         if (!sugov_update_next_freq(sg_policy, time, next_freq))
 127                 return;
 128
 129         if (!sg_policy->work_in_progress) {
 130                 sg_policy->work_in_progress = true;
 131                 irq_work_queue(&sg_policy->irq_work);
 132         }
 133 }
 134
 135 /**
 136  * get_next_freq - Compute a new frequency for a given cpufreq policy.
 137  * @sg_policy: schedutil policy object to compute the new frequency for.
 138  * @util: Current CPU utilization.
 139  * @max: CPU capacity.
 140  *
 141  * If the utilization is frequency-invariant, choose the new frequency to be
 142  * proportional to it, that is
 143  *
 144  * next_freq = C * max_freq * util / max
 145  *
 146  * Otherwise, approximate the would-be frequency-invariant utilization by
 147  * util_raw * (curr_freq / max_freq) which leads to
 148  *
 149  * next_freq = C * curr_freq * util_raw / max
 150  *
 151  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
 152  *
 153  * The lowest driver-supported frequency which is equal or greater than the raw
 154  * next_freq (as calculated above) is returned, subject to policy min/max and
 155  * cpufreq driver limitations.
 156  */
 157 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
 158                                   unsigned long util, unsigned long max)
 159 {
 160         struct cpufreq_policy *policy = sg_policy->policy;
 161         unsigned int freq = arch_scale_freq_invariant() ?
 162                                 policy->cpuinfo.max_freq : policy->cur;
 163
 164         freq = map_util_freq(util, freq, max);
 165
 166         if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
 167                 return sg_policy->next_freq;
 168
 169         sg_policy->cached_raw_freq = freq;
 170         return cpufreq_driver_resolve_freq(policy, freq);
 171 }
 172
 173 /*
 174  * This function computes an effective utilization for the given CPU, to be
 175  * used for frequency selection given the linear relation: f = u * f_max.
 176  *
 177  * The scheduler tracks the following metrics:
 178  *
 179  *   cpu_util_{cfs,rt,dl,irq}()
 180  *   cpu_bw_dl()
 181  *
 182  * Where the cfs,rt and dl util numbers are tracked with the same metric and
 183  * synchronized windows and are thus directly comparable.
 184  *
 185  * The cfs,rt,dl utilization are the running times measured with rq->clock_task
 186  * which excludes things like IRQ and steal-time. These latter are then accrued
 187  * in the irq utilization.
 188  *
 189  * The DL bandwidth number otoh is not a measured metric but a value computed
 190  * based on the task model parameters and gives the minimal utilization
 191  * required to meet deadlines.
 192  */
 193 unsigned long schedutil_cpu_util(int cpu, unsigned long util_cfs,
 194                                  unsigned long max, enum schedutil_type type,
 195                                  struct task_struct *p)
 196 {
 197         unsigned long dl_util, util, irq;
 198         struct rq *rq = cpu_rq(cpu);
 199
 200         if (!uclamp_is_used() &&
 201             type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
 202                 return max;
 203         }
 204
 205         /*
 206          * Early check to see if IRQ/steal time saturates the CPU, can be
 207          * because of inaccuracies in how we track these -- see
 208          * update_irq_load_avg().
 209          */
 210         irq = cpu_util_irq(rq);
 211         if (unlikely(irq >= max))
 212                 return max;
 213
 214         /*
 215          * Because the time spend on RT/DL tasks is visible as 'lost' time to
 216          * CFS tasks and we use the same metric to track the effective
 217          * utilization (PELT windows are synchronized) we can directly add them
 218          * to obtain the CPU's actual utilization.
 219          *
 220          * CFS and RT utilization can be boosted or capped, depending on
 221          * utilization clamp constraints requested by currently RUNNABLE
 222          * tasks.
 223          * When there are no CFS RUNNABLE tasks, clamps are released and
 224          * frequency will be gracefully reduced with the utilization decay.
 225          */
 226         util = util_cfs + cpu_util_rt(rq);
 227         if (type == FREQUENCY_UTIL)
 228                 util = uclamp_rq_util_with(rq, util, p);
 229
 230         dl_util = cpu_util_dl(rq);
 231
 232         /*
 233          * For frequency selection we do not make cpu_util_dl() a permanent part
 234          * of this sum because we want to use cpu_bw_dl() later on, but we need
 235          * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
 236          * that we select f_max when there is no idle time.
 237          *
 238          * NOTE: numerical errors or stop class might cause us to not quite hit
 239          * saturation when we should -- something for later.
 240          */
 241         if (util + dl_util >= max)
 242                 return max;
 243
 244         /*
 245          * OTOH, for energy computation we need the estimated running time, so
 246          * include util_dl and ignore dl_bw.
 247          */
 248         if (type == ENERGY_UTIL)
 249                 util += dl_util;
 250
 251         /*
 252          * There is still idle time; further improve the number by using the
 253          * irq metric. Because IRQ/steal time is hidden from the task clock we
 254          * need to scale the task numbers:
 255          *
 256          *              max - irq
 257          *   U' = irq + --------- * U
 258          *                 max
 259          */
 260         util = scale_irq_capacity(util, irq, max);
 261         util += irq;
 262
 263         /*
 264          * Bandwidth required by DEADLINE must always be granted while, for
 265          * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
 266          * to gracefully reduce the frequency when no tasks show up for longer
 267          * periods of time.
 268          *
 269          * Ideally we would like to set bw_dl as min/guaranteed freq and util +
 270          * bw_dl as requested freq. However, cpufreq is not yet ready for such
 271          * an interface. So, we only do the latter for now.
 272          */
 273         if (type == FREQUENCY_UTIL)
 274                 util += cpu_bw_dl(rq);
 275
 276         return min(max, util);
 277 }
 278
 279 static unsigned long sugov_get_util(struct sugov_cpu *sg_cpu)
 280 {
 281         struct rq *rq = cpu_rq(sg_cpu->cpu);
 282         unsigned long util = cpu_util_cfs(rq);
 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
 288         return schedutil_cpu_util(sg_cpu->cpu, util, max, 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  * @util: the utilization to (eventually) boost
 366  * @max: the maximum value the utilization can be boosted to
 367  *
 368  * A CPU running a task which woken up after an IO operation can have its
 369  * utilization boosted to speed up the completion of those IO operations.
 370  * The IO boost value is increased each time a task wakes up from IO, in
 371  * sugov_iowait_apply(), and it's instead decreased by this function,
 372  * each time an increase has not been requested (!iowait_boost_pending).
 373  *
 374  * A CPU which also appears to have been idle for at least one tick has also
 375  * its IO boost utilization reset.
 376  *
 377  * This mechanism is designed to boost high frequently IO waiting tasks, while
 378  * being more conservative on tasks which does sporadic IO operations.
 379  */
 380 static unsigned long sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
 381                                         unsigned long util, unsigned long max)
 382 {
 383         unsigned long boost;
 384
 385         /* No boost currently required */
 386         if (!sg_cpu->iowait_boost)
 387                 return util;
 388
 389         /* Reset boost if the CPU appears to have been idle enough */
 390         if (sugov_iowait_reset(sg_cpu, time, false))
 391                 return util;
 392
 393         if (!sg_cpu->iowait_boost_pending) {
 394                 /*
 395                  * No boost pending; reduce the boost value.
 396                  */
 397                 sg_cpu->iowait_boost >>= 1;
 398                 if (sg_cpu->iowait_boost < IOWAIT_BOOST_MIN) {
 399                         sg_cpu->iowait_boost = 0;
 400                         return util;
 401                 }
 402         }
 403
 404         sg_cpu->iowait_boost_pending = false;
 405
 406         /*
 407          * @util is already in capacity scale; convert iowait_boost
 408          * into the same scale so we can compare.
 409          */
 410         boost = (sg_cpu->iowait_boost * max) >> SCHED_CAPACITY_SHIFT;
 411         return max(boost, util);
 412 }
 413
 414 #ifdef CONFIG_NO_HZ_COMMON
 415 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
 416 {
 417         unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
 418         bool ret = idle_calls == sg_cpu->saved_idle_calls;
 419
 420         sg_cpu->saved_idle_calls = idle_calls;
 421         return ret;
 422 }
 423 #else
 424 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
 425 #endif /* CONFIG_NO_HZ_COMMON */
 426
 427 /*
 428  * Make sugov_should_update_freq() ignore the rate limit when DL
 429  * has increased the utilization.
 430  */
 431 static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu, struct sugov_policy *sg_policy)
 432 {
 433         if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_dl)
 434                 sg_policy->limits_changed = true;
 435 }
 436
 437 static void sugov_update_single(struct update_util_data *hook, u64 time,
 438                                 unsigned int flags)
 439 {
 440         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 441         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 442         unsigned long util, max;
 443         unsigned int next_f;
 444         unsigned int cached_freq = sg_policy->cached_raw_freq;
 445
 446         sugov_iowait_boost(sg_cpu, time, flags);
 447         sg_cpu->last_update = time;
 448
 449         ignore_dl_rate_limit(sg_cpu, sg_policy);
 450
 451         if (!sugov_should_update_freq(sg_policy, time))
 452                 return;
 453
 454         util = sugov_get_util(sg_cpu);
 455         max = sg_cpu->max;
 456         util = sugov_iowait_apply(sg_cpu, time, util, max);
 457         next_f = get_next_freq(sg_policy, util, max);
 458         /*
 459          * Do not reduce the frequency if the CPU has not been idle
 460          * recently, as the reduction is likely to be premature then.
 461          */
 462         if (sugov_cpu_is_busy(sg_cpu) && next_f < sg_policy->next_freq) {
 463                 next_f = sg_policy->next_freq;
 464
 465                 /* Restore cached freq as next_freq has changed */
 466                 sg_policy->cached_raw_freq = cached_freq;
 467         }
 468
 469         /*
 470          * This code runs under rq->lock for the target CPU, so it won't run
 471          * concurrently on two different CPUs for the same target and it is not
 472          * necessary to acquire the lock in the fast switch case.
 473          */
 474         if (sg_policy->policy->fast_switch_enabled) {
 475                 sugov_fast_switch(sg_policy, time, next_f);
 476         } else {
 477                 raw_spin_lock(&sg_policy->update_lock);
 478                 sugov_deferred_update(sg_policy, time, next_f);
 479                 raw_spin_unlock(&sg_policy->update_lock);
 480         }
 481 }
 482
 483 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
 484 {
 485         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 486         struct cpufreq_policy *policy = sg_policy->policy;
 487         unsigned long util = 0, max = 1;
 488         unsigned int j;
 489
 490         for_each_cpu(j, policy->cpus) {
 491                 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
 492                 unsigned long j_util, j_max;
 493
 494                 j_util = sugov_get_util(j_sg_cpu);
 495                 j_max = j_sg_cpu->max;
 496                 j_util = sugov_iowait_apply(j_sg_cpu, time, j_util, j_max);
 497
 498                 if (j_util * max > j_max * util) {
 499                         util = j_util;
 500                         max = j_max;
 501                 }
 502         }
 503
 504         return get_next_freq(sg_policy, util, max);
 505 }
 506
 507 static void
 508 sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
 509 {
 510         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 511         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 512         unsigned int next_f;
 513
 514         raw_spin_lock(&sg_policy->update_lock);
 515
 516         sugov_iowait_boost(sg_cpu, time, flags);
 517         sg_cpu->last_update = time;
 518
 519         ignore_dl_rate_limit(sg_cpu, sg_policy);
 520
 521         if (sugov_should_update_freq(sg_policy, time)) {
 522                 next_f = sugov_next_freq_shared(sg_cpu, time);
 523
 524                 if (sg_policy->policy->fast_switch_enabled)
 525                         sugov_fast_switch(sg_policy, time, next_f);
 526                 else
 527                         sugov_deferred_update(sg_policy, time, next_f);
 528         }
 529
 530         raw_spin_unlock(&sg_policy->update_lock);
 531 }
 532
 533 static void sugov_work(struct kthread_work *work)
 534 {
 535         struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
 536         unsigned int freq;
 537         unsigned long flags;
 538
 539         /*
 540          * Hold sg_policy->update_lock shortly to handle the case where:
 541          * incase sg_policy->next_freq is read here, and then updated by
 542          * sugov_deferred_update() just before work_in_progress is set to false
 543          * here, we may miss queueing the new update.
 544          *
 545          * Note: If a work was queued after the update_lock is released,
 546          * sugov_work() will just be called again by kthread_work code; and the
 547          * request will be proceed before the sugov thread sleeps.
 548          */
 549         raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
 550         freq = sg_policy->next_freq;
 551         sg_policy->work_in_progress = false;
 552         raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
 553
 554         mutex_lock(&sg_policy->work_lock);
 555         __cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
 556         mutex_unlock(&sg_policy->work_lock);
 557 }
 558
 559 static void sugov_irq_work(struct irq_work *irq_work)
 560 {
 561         struct sugov_policy *sg_policy;
 562
 563         sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
 564
 565         kthread_queue_work(&sg_policy->worker, &sg_policy->work);
 566 }
 567
 568 /************************** sysfs interface ************************/
 569
 570 static struct sugov_tunables *global_tunables;
 571 static DEFINE_MUTEX(global_tunables_lock);
 572
 573 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
 574 {
 575         return container_of(attr_set, struct sugov_tunables, attr_set);
 576 }
 577
 578 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
 579 {
 580         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 581
 582         return sprintf(buf, "%u\n", tunables->rate_limit_us);
 583 }
 584
 585 static ssize_t
 586 rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
 587 {
 588         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 589         struct sugov_policy *sg_policy;
 590         unsigned int rate_limit_us;
 591
 592         if (kstrtouint(buf, 10, &rate_limit_us))
 593                 return -EINVAL;
 594
 595         tunables->rate_limit_us = rate_limit_us;
 596
 597         list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
 598                 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
 599
 600         return count;
 601 }
 602
 603 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
 604
 605 static struct attribute *sugov_attrs[] = {
 606         &rate_limit_us.attr,
 607         NULL
 608 };
 609 ATTRIBUTE_GROUPS(sugov);
 610
 611 static struct kobj_type sugov_tunables_ktype = {
 612         .default_groups = sugov_groups,
 613         .sysfs_ops = &governor_sysfs_ops,
 614 };
 615
 616 /********************** cpufreq governor interface *********************/
 617
 618 struct cpufreq_governor schedutil_gov;
 619
 620 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
 621 {
 622         struct sugov_policy *sg_policy;
 623
 624         sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
 625         if (!sg_policy)
 626                 return NULL;
 627
 628         sg_policy->policy = policy;
 629         raw_spin_lock_init(&sg_policy->update_lock);
 630         return sg_policy;
 631 }
 632
 633 static void sugov_policy_free(struct sugov_policy *sg_policy)
 634 {
 635         kfree(sg_policy);
 636 }
 637
 638 static int sugov_kthread_create(struct sugov_policy *sg_policy)
 639 {
 640         struct task_struct *thread;
 641         struct sched_attr attr = {
 642                 .size           = sizeof(struct sched_attr),
 643                 .sched_policy   = SCHED_DEADLINE,
 644                 .sched_flags    = SCHED_FLAG_SUGOV,
 645                 .sched_nice     = 0,
 646                 .sched_priority = 0,
 647                 /*
 648                  * Fake (unused) bandwidth; workaround to "fix"
 649                  * priority inheritance.
 650                  */
 651                 .sched_runtime  =  1000000,
 652                 .sched_deadline = 10000000,
 653                 .sched_period   = 10000000,
 654         };
 655         struct cpufreq_policy *policy = sg_policy->policy;
 656         int ret;
 657
 658         /* kthread only required for slow path */
 659         if (policy->fast_switch_enabled)
 660                 return 0;
 661
 662         kthread_init_work(&sg_policy->work, sugov_work);
 663         kthread_init_worker(&sg_policy->worker);
 664         thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
 665                                 "sugov:%d",
 666                                 cpumask_first(policy->related_cpus));
 667         if (IS_ERR(thread)) {
 668                 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
 669                 return PTR_ERR(thread);
 670         }
 671
 672         ret = sched_setattr_nocheck(thread, &attr);
 673         if (ret) {
 674                 kthread_stop(thread);
 675                 pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
 676                 return ret;
 677         }
 678
 679         sg_policy->thread = thread;
 680         kthread_bind_mask(thread, policy->related_cpus);
 681         init_irq_work(&sg_policy->irq_work, sugov_irq_work);
 682         mutex_init(&sg_policy->work_lock);
 683
 684         wake_up_process(thread);
 685
 686         return 0;
 687 }
 688
 689 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
 690 {
 691         /* kthread only required for slow path */
 692         if (sg_policy->policy->fast_switch_enabled)
 693                 return;
 694
 695         kthread_flush_worker(&sg_policy->worker);
 696         kthread_stop(sg_policy->thread);
 697         mutex_destroy(&sg_policy->work_lock);
 698 }
 699
 700 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
 701 {
 702         struct sugov_tunables *tunables;
 703
 704         tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
 705         if (tunables) {
 706                 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
 707                 if (!have_governor_per_policy())
 708                         global_tunables = tunables;
 709         }
 710         return tunables;
 711 }
 712
 713 static void sugov_tunables_free(struct sugov_tunables *tunables)
 714 {
 715         if (!have_governor_per_policy())
 716                 global_tunables = NULL;
 717
 718         kfree(tunables);
 719 }
 720
 721 static int sugov_init(struct cpufreq_policy *policy)
 722 {
 723         struct sugov_policy *sg_policy;
 724         struct sugov_tunables *tunables;
 725         int ret = 0;
 726
 727         /* State should be equivalent to EXIT */
 728         if (policy->governor_data)
 729                 return -EBUSY;
 730
 731         cpufreq_enable_fast_switch(policy);
 732
 733         sg_policy = sugov_policy_alloc(policy);
 734         if (!sg_policy) {
 735                 ret = -ENOMEM;
 736                 goto disable_fast_switch;
 737         }
 738
 739         ret = sugov_kthread_create(sg_policy);
 740         if (ret)
 741                 goto free_sg_policy;
 742
 743         mutex_lock(&global_tunables_lock);
 744
 745         if (global_tunables) {
 746                 if (WARN_ON(have_governor_per_policy())) {
 747                         ret = -EINVAL;
 748                         goto stop_kthread;
 749                 }
 750                 policy->governor_data = sg_policy;
 751                 sg_policy->tunables = global_tunables;
 752
 753                 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
 754                 goto out;
 755         }
 756
 757         tunables = sugov_tunables_alloc(sg_policy);
 758         if (!tunables) {
 759                 ret = -ENOMEM;
 760                 goto stop_kthread;
 761         }
 762
 763         tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
 764
 765         policy->governor_data = sg_policy;
 766         sg_policy->tunables = tunables;
 767
 768         ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
 769                                    get_governor_parent_kobj(policy), "%s",
 770                                    schedutil_gov.name);
 771         if (ret)
 772                 goto fail;
 773
 774 out:
 775         mutex_unlock(&global_tunables_lock);
 776         return 0;
 777
 778 fail:
 779         kobject_put(&tunables->attr_set.kobj);
 780         policy->governor_data = NULL;
 781         sugov_tunables_free(tunables);
 782
 783 stop_kthread:
 784         sugov_kthread_stop(sg_policy);
 785         mutex_unlock(&global_tunables_lock);
 786
 787 free_sg_policy:
 788         sugov_policy_free(sg_policy);
 789
 790 disable_fast_switch:
 791         cpufreq_disable_fast_switch(policy);
 792
 793         pr_err("initialization failed (error %d)\n", ret);
 794         return ret;
 795 }
 796
 797 static void sugov_exit(struct cpufreq_policy *policy)
 798 {
 799         struct sugov_policy *sg_policy = policy->governor_data;
 800         struct sugov_tunables *tunables = sg_policy->tunables;
 801         unsigned int count;
 802
 803         mutex_lock(&global_tunables_lock);
 804
 805         count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
 806         policy->governor_data = NULL;
 807         if (!count)
 808                 sugov_tunables_free(tunables);
 809
 810         mutex_unlock(&global_tunables_lock);
 811
 812         sugov_kthread_stop(sg_policy);
 813         sugov_policy_free(sg_policy);
 814         cpufreq_disable_fast_switch(policy);
 815 }
 816
 817 static int sugov_start(struct cpufreq_policy *policy)
 818 {
 819         struct sugov_policy *sg_policy = policy->governor_data;
 820         unsigned int cpu;
 821
 822         sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
 823         sg_policy->last_freq_update_time        = 0;
 824         sg_policy->next_freq                    = 0;
 825         sg_policy->work_in_progress             = false;
 826         sg_policy->limits_changed               = false;
 827         sg_policy->cached_raw_freq              = 0;
 828
 829         sg_policy->need_freq_update = cpufreq_driver_test_flags(CPUFREQ_NEED_UPDATE_LIMITS);
 830
 831         for_each_cpu(cpu, policy->cpus) {
 832                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 833
 834                 memset(sg_cpu, 0, sizeof(*sg_cpu));
 835                 sg_cpu->cpu                     = cpu;
 836                 sg_cpu->sg_policy               = sg_policy;
 837         }
 838
 839         for_each_cpu(cpu, policy->cpus) {
 840                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 841
 842                 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
 843                                              policy_is_shared(policy) ?
 844                                                         sugov_update_shared :
 845                                                         sugov_update_single);
 846         }
 847         return 0;
 848 }
 849
 850 static void sugov_stop(struct cpufreq_policy *policy)
 851 {
 852         struct sugov_policy *sg_policy = policy->governor_data;
 853         unsigned int cpu;
 854
 855         for_each_cpu(cpu, policy->cpus)
 856                 cpufreq_remove_update_util_hook(cpu);
 857
 858         synchronize_rcu();
 859
 860         if (!policy->fast_switch_enabled) {
 861                 irq_work_sync(&sg_policy->irq_work);
 862                 kthread_cancel_work_sync(&sg_policy->work);
 863         }
 864 }
 865
 866 static void sugov_limits(struct cpufreq_policy *policy)
 867 {
 868         struct sugov_policy *sg_policy = policy->governor_data;
 869
 870         if (!policy->fast_switch_enabled) {
 871                 mutex_lock(&sg_policy->work_lock);
 872                 cpufreq_policy_apply_limits(policy);
 873                 mutex_unlock(&sg_policy->work_lock);
 874         }
 875
 876         sg_policy->limits_changed = true;
 877 }
 878
 879 struct cpufreq_governor schedutil_gov = {
 880         .name                   = "schedutil",
 881         .owner                  = THIS_MODULE,
 882         .flags                  = CPUFREQ_GOV_DYNAMIC_SWITCHING,
 883         .init                   = sugov_init,
 884         .exit                   = sugov_exit,
 885         .start                  = sugov_start,
 886         .stop                   = sugov_stop,
 887         .limits                 = sugov_limits,
 888 };
 889
 890 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
 891 struct cpufreq_governor *cpufreq_default_governor(void)
 892 {
 893         return &schedutil_gov;
 894 }
 895 #endif
 896
 897 cpufreq_governor_init(schedutil_gov);
 898
 899 #ifdef CONFIG_ENERGY_MODEL
 900 static void rebuild_sd_workfn(struct work_struct *work)
 901 {
 902         rebuild_sched_domains_energy();
 903 }
 904 static DECLARE_WORK(rebuild_sd_work, rebuild_sd_workfn);
 905
 906 /*
 907  * EAS shouldn't be attempted without sugov, so rebuild the sched_domains
 908  * on governor changes to make sure the scheduler knows about it.
 909  */
 910 void sched_cpufreq_governor_change(struct cpufreq_policy *policy,
 911                                   struct cpufreq_governor *old_gov)
 912 {
 913         if (old_gov == &schedutil_gov || policy->governor == &schedutil_gov) {
 914                 /*
 915                  * When called from the cpufreq_register_driver() path, the
 916                  * cpu_hotplug_lock is already held, so use a work item to
 917                  * avoid nested locking in rebuild_sched_domains().
 918                  */
 919                 schedule_work(&rebuild_sd_work);
 920         }
 921
 922 }
 923 #endif