kernel/sched/cpufreq_schedutil.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * CPUFreq governor based on scheduler-provided CPU utilization data.
   4  *
   5  * Copyright (C) 2016, Intel Corporation
   6  * Author: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
   7  */
   8
   9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  11 #include "sched.h"
  12
  13 #include <linux/sched/cpufreq.h>
  14 #include <trace/events/power.h>
  15
  16 #define IOWAIT_BOOST_MIN        (SCHED_CAPACITY_SCALE / 8)
  17
  18 struct sugov_tunables {
  19         struct gov_attr_set     attr_set;
  20         unsigned int            rate_limit_us;
  21 };
  22
  23 struct sugov_policy {
  24         struct cpufreq_policy   *policy;
  25
  26         struct sugov_tunables   *tunables;
  27         struct list_head        tunables_hook;
  28
  29         raw_spinlock_t          update_lock;    /* For shared policies */
  30         u64                     last_freq_update_time;
  31         s64                     freq_update_delay_ns;
  32         unsigned int            next_freq;
  33         unsigned int            cached_raw_freq;
  34
  35         /* The next fields are only needed if fast switch cannot be used: */
  36         struct                  irq_work irq_work;
  37         struct                  kthread_work work;
  38         struct                  mutex work_lock;
  39         struct                  kthread_worker worker;
  40         struct task_struct      *thread;
  41         bool                    work_in_progress;
  42
  43         bool                    need_freq_update;
  44 };
  45
  46 struct sugov_cpu {
  47         struct update_util_data update_util;
  48         struct sugov_policy     *sg_policy;
  49         unsigned int            cpu;
  50
  51         bool                    iowait_boost_pending;
  52         unsigned int            iowait_boost;
  53         u64                     last_update;
  54
  55         unsigned long           bw_dl;
  56         unsigned long           max;
  57
  58         /* The field below is for single-CPU policies only: */
  59 #ifdef CONFIG_NO_HZ_COMMON
  60         unsigned long           saved_idle_calls;
  61 #endif
  62 };
  63
  64 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
  65
  66 /************************ Governor internals ***********************/
  67
  68 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
  69 {
  70         s64 delta_ns;
  71
  72         /*
  73          * Since cpufreq_update_util() is called with rq->lock held for
  74          * the @target_cpu, our per-CPU data is fully serialized.
  75          *
  76          * However, drivers cannot in general deal with cross-CPU
  77          * requests, so while get_next_freq() will work, our
  78          * sugov_update_commit() call may not for the fast switching platforms.
  79          *
  80          * Hence stop here for remote requests if they aren't supported
  81          * by the hardware, as calculating the frequency is pointless if
  82          * we cannot in fact act on it.
  83          *
  84          * For the slow switching platforms, the kthread is always scheduled on
  85          * the right set of CPUs and any CPU can find the next frequency and
  86          * schedule the kthread.
  87          */
  88         if (sg_policy->policy->fast_switch_enabled &&
  89             !cpufreq_this_cpu_can_update(sg_policy->policy))
  90                 return false;
  91
  92         if (unlikely(sg_policy->need_freq_update))
  93                 return true;
  94
  95         delta_ns = time - sg_policy->last_freq_update_time;
  96
  97         return delta_ns >= sg_policy->freq_update_delay_ns;
  98 }
  99
 100 static bool sugov_update_next_freq(struct sugov_policy *sg_policy, u64 time,
 101                                    unsigned int next_freq)
 102 {
 103         if (sg_policy->next_freq == next_freq)
 104                 return false;
 105
 106         sg_policy->next_freq = next_freq;
 107         sg_policy->last_freq_update_time = time;
 108
 109         return true;
 110 }
 111
 112 static void sugov_fast_switch(struct sugov_policy *sg_policy, u64 time,
 113                               unsigned int next_freq)
 114 {
 115         struct cpufreq_policy *policy = sg_policy->policy;
 116
 117         if (!sugov_update_next_freq(sg_policy, time, next_freq))
 118                 return;
 119
 120         next_freq = cpufreq_driver_fast_switch(policy, next_freq);
 121         if (!next_freq)
 122                 return;
 123
 124         policy->cur = next_freq;
 125         trace_cpu_frequency(next_freq, smp_processor_id());
 126 }
 127
 128 static void sugov_deferred_update(struct sugov_policy *sg_policy, u64 time,
 129                                   unsigned int next_freq)
 130 {
 131         if (!sugov_update_next_freq(sg_policy, time, next_freq))
 132                 return;
 133
 134         if (!sg_policy->work_in_progress) {
 135                 sg_policy->work_in_progress = true;
 136                 irq_work_queue(&sg_policy->irq_work);
 137         }
 138 }
 139
 140 /**
 141  * get_next_freq - Compute a new frequency for a given cpufreq policy.
 142  * @sg_policy: schedutil policy object to compute the new frequency for.
 143  * @util: Current CPU utilization.
 144  * @max: CPU capacity.
 145  *
 146  * If the utilization is frequency-invariant, choose the new frequency to be
 147  * proportional to it, that is
 148  *
 149  * next_freq = C * max_freq * util / max
 150  *
 151  * Otherwise, approximate the would-be frequency-invariant utilization by
 152  * util_raw * (curr_freq / max_freq) which leads to
 153  *
 154  * next_freq = C * curr_freq * util_raw / max
 155  *
 156  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
 157  *
 158  * The lowest driver-supported frequency which is equal or greater than the raw
 159  * next_freq (as calculated above) is returned, subject to policy min/max and
 160  * cpufreq driver limitations.
 161  */
 162 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
 163                                   unsigned long util, unsigned long max)
 164 {
 165         struct cpufreq_policy *policy = sg_policy->policy;
 166         unsigned int freq = arch_scale_freq_invariant() ?
 167                                 policy->cpuinfo.max_freq : policy->cur;
 168
 169         freq = map_util_freq(util, freq, max);
 170
 171         if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
 172                 return sg_policy->next_freq;
 173
 174         sg_policy->need_freq_update = false;
 175         sg_policy->cached_raw_freq = freq;
 176         return cpufreq_driver_resolve_freq(policy, freq);
 177 }
 178
 179 /*
 180  * This function computes an effective utilization for the given CPU, to be
 181  * used for frequency selection given the linear relation: f = u * f_max.
 182  *
 183  * The scheduler tracks the following metrics:
 184  *
 185  *   cpu_util_{cfs,rt,dl,irq}()
 186  *   cpu_bw_dl()
 187  *
 188  * Where the cfs,rt and dl util numbers are tracked with the same metric and
 189  * synchronized windows and are thus directly comparable.
 190  *
 191  * The cfs,rt,dl utilization are the running times measured with rq->clock_task
 192  * which excludes things like IRQ and steal-time. These latter are then accrued
 193  * in the irq utilization.
 194  *
 195  * The DL bandwidth number otoh is not a measured metric but a value computed
 196  * based on the task model parameters and gives the minimal utilization
 197  * required to meet deadlines.
 198  */
 199 unsigned long schedutil_freq_util(int cpu, unsigned long util_cfs,
 200                                   unsigned long max, enum schedutil_type type)
 201 {
 202         unsigned long dl_util, util, irq;
 203         struct rq *rq = cpu_rq(cpu);
 204
 205         if (type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt))
 206                 return max;
 207
 208         /*
 209          * Early check to see if IRQ/steal time saturates the CPU, can be
 210          * because of inaccuracies in how we track these -- see
 211          * update_irq_load_avg().
 212          */
 213         irq = cpu_util_irq(rq);
 214         if (unlikely(irq >= max))
 215                 return max;
 216
 217         /*
 218          * Because the time spend on RT/DL tasks is visible as 'lost' time to
 219          * CFS tasks and we use the same metric to track the effective
 220          * utilization (PELT windows are synchronized) we can directly add them
 221          * to obtain the CPU's actual utilization.
 222          */
 223         util = util_cfs;
 224         util += cpu_util_rt(rq);
 225
 226         dl_util = cpu_util_dl(rq);
 227
 228         /*
 229          * For frequency selection we do not make cpu_util_dl() a permanent part
 230          * of this sum because we want to use cpu_bw_dl() later on, but we need
 231          * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
 232          * that we select f_max when there is no idle time.
 233          *
 234          * NOTE: numerical errors or stop class might cause us to not quite hit
 235          * saturation when we should -- something for later.
 236          */
 237         if (util + dl_util >= max)
 238                 return max;
 239
 240         /*
 241          * OTOH, for energy computation we need the estimated running time, so
 242          * include util_dl and ignore dl_bw.
 243          */
 244         if (type == ENERGY_UTIL)
 245                 util += dl_util;
 246
 247         /*
 248          * There is still idle time; further improve the number by using the
 249          * irq metric. Because IRQ/steal time is hidden from the task clock we
 250          * need to scale the task numbers:
 251          *
 252          *              1 - irq
 253          *   U' = irq + ------- * U
 254          *                max
 255          */
 256         util = scale_irq_capacity(util, irq, max);
 257         util += irq;
 258
 259         /*
 260          * Bandwidth required by DEADLINE must always be granted while, for
 261          * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
 262          * to gracefully reduce the frequency when no tasks show up for longer
 263          * periods of time.
 264          *
 265          * Ideally we would like to set bw_dl as min/guaranteed freq and util +
 266          * bw_dl as requested freq. However, cpufreq is not yet ready for such
 267          * an interface. So, we only do the latter for now.
 268          */
 269         if (type == FREQUENCY_UTIL)
 270                 util += cpu_bw_dl(rq);
 271
 272         return min(max, util);
 273 }
 274
 275 static unsigned long sugov_get_util(struct sugov_cpu *sg_cpu)
 276 {
 277         struct rq *rq = cpu_rq(sg_cpu->cpu);
 278         unsigned long util = cpu_util_cfs(rq);
 279         unsigned long max = arch_scale_cpu_capacity(NULL, sg_cpu->cpu);
 280
 281         sg_cpu->max = max;
 282         sg_cpu->bw_dl = cpu_bw_dl(rq);
 283
 284         return schedutil_freq_util(sg_cpu->cpu, util, max, FREQUENCY_UTIL);
 285 }
 286
 287 /**
 288  * sugov_iowait_reset() - Reset the IO boost status of a CPU.
 289  * @sg_cpu: the sugov data for the CPU to boost
 290  * @time: the update time from the caller
 291  * @set_iowait_boost: true if an IO boost has been requested
 292  *
 293  * The IO wait boost of a task is disabled after a tick since the last update
 294  * of a CPU. If a new IO wait boost is requested after more then a tick, then
 295  * we enable the boost starting from IOWAIT_BOOST_MIN, which improves energy
 296  * efficiency by ignoring sporadic wakeups from IO.
 297  */
 298 static bool sugov_iowait_reset(struct sugov_cpu *sg_cpu, u64 time,
 299                                bool set_iowait_boost)
 300 {
 301         s64 delta_ns = time - sg_cpu->last_update;
 302
 303         /* Reset boost only if a tick has elapsed since last request */
 304         if (delta_ns <= TICK_NSEC)
 305                 return false;
 306
 307         sg_cpu->iowait_boost = set_iowait_boost ? IOWAIT_BOOST_MIN : 0;
 308         sg_cpu->iowait_boost_pending = set_iowait_boost;
 309
 310         return true;
 311 }
 312
 313 /**
 314  * sugov_iowait_boost() - Updates the IO boost status of a CPU.
 315  * @sg_cpu: the sugov data for the CPU to boost
 316  * @time: the update time from the caller
 317  * @flags: SCHED_CPUFREQ_IOWAIT if the task is waking up after an IO wait
 318  *
 319  * Each time a task wakes up after an IO operation, the CPU utilization can be
 320  * boosted to a certain utilization which doubles at each "frequent and
 321  * successive" wakeup from IO, ranging from IOWAIT_BOOST_MIN to the utilization
 322  * of the maximum OPP.
 323  *
 324  * To keep doubling, an IO boost has to be requested at least once per tick,
 325  * otherwise we restart from the utilization of the minimum OPP.
 326  */
 327 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
 328                                unsigned int flags)
 329 {
 330         bool set_iowait_boost = flags & SCHED_CPUFREQ_IOWAIT;
 331
 332         /* Reset boost if the CPU appears to have been idle enough */
 333         if (sg_cpu->iowait_boost &&
 334             sugov_iowait_reset(sg_cpu, time, set_iowait_boost))
 335                 return;
 336
 337         /* Boost only tasks waking up after IO */
 338         if (!set_iowait_boost)
 339                 return;
 340
 341         /* Ensure boost doubles only one time at each request */
 342         if (sg_cpu->iowait_boost_pending)
 343                 return;
 344         sg_cpu->iowait_boost_pending = true;
 345
 346         /* Double the boost at each request */
 347         if (sg_cpu->iowait_boost) {
 348                 sg_cpu->iowait_boost =
 349                         min_t(unsigned int, sg_cpu->iowait_boost << 1, SCHED_CAPACITY_SCALE);
 350                 return;
 351         }
 352
 353         /* First wakeup after IO: start with minimum boost */
 354         sg_cpu->iowait_boost = IOWAIT_BOOST_MIN;
 355 }
 356
 357 /**
 358  * sugov_iowait_apply() - Apply the IO boost to a CPU.
 359  * @sg_cpu: the sugov data for the cpu to boost
 360  * @time: the update time from the caller
 361  * @util: the utilization to (eventually) boost
 362  * @max: the maximum value the utilization can be boosted to
 363  *
 364  * A CPU running a task which woken up after an IO operation can have its
 365  * utilization boosted to speed up the completion of those IO operations.
 366  * The IO boost value is increased each time a task wakes up from IO, in
 367  * sugov_iowait_apply(), and it's instead decreased by this function,
 368  * each time an increase has not been requested (!iowait_boost_pending).
 369  *
 370  * A CPU which also appears to have been idle for at least one tick has also
 371  * its IO boost utilization reset.
 372  *
 373  * This mechanism is designed to boost high frequently IO waiting tasks, while
 374  * being more conservative on tasks which does sporadic IO operations.
 375  */
 376 static unsigned long sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
 377                                         unsigned long util, unsigned long max)
 378 {
 379         unsigned long boost;
 380
 381         /* No boost currently required */
 382         if (!sg_cpu->iowait_boost)
 383                 return util;
 384
 385         /* Reset boost if the CPU appears to have been idle enough */
 386         if (sugov_iowait_reset(sg_cpu, time, false))
 387                 return util;
 388
 389         if (!sg_cpu->iowait_boost_pending) {
 390                 /*
 391                  * No boost pending; reduce the boost value.
 392                  */
 393                 sg_cpu->iowait_boost >>= 1;
 394                 if (sg_cpu->iowait_boost < IOWAIT_BOOST_MIN) {
 395                         sg_cpu->iowait_boost = 0;
 396                         return util;
 397                 }
 398         }
 399
 400         sg_cpu->iowait_boost_pending = false;
 401
 402         /*
 403          * @util is already in capacity scale; convert iowait_boost
 404          * into the same scale so we can compare.
 405          */
 406         boost = (sg_cpu->iowait_boost * max) >> SCHED_CAPACITY_SHIFT;
 407         return max(boost, util);
 408 }
 409
 410 #ifdef CONFIG_NO_HZ_COMMON
 411 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
 412 {
 413         unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
 414         bool ret = idle_calls == sg_cpu->saved_idle_calls;
 415
 416         sg_cpu->saved_idle_calls = idle_calls;
 417         return ret;
 418 }
 419 #else
 420 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
 421 #endif /* CONFIG_NO_HZ_COMMON */
 422
 423 /*
 424  * Make sugov_should_update_freq() ignore the rate limit when DL
 425  * has increased the utilization.
 426  */
 427 static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu, struct sugov_policy *sg_policy)
 428 {
 429         if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_dl)
 430                 sg_policy->need_freq_update = true;
 431 }
 432
 433 static void sugov_update_single(struct update_util_data *hook, u64 time,
 434                                 unsigned int flags)
 435 {
 436         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 437         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 438         unsigned long util, max;
 439         unsigned int next_f;
 440         bool busy;
 441
 442         sugov_iowait_boost(sg_cpu, time, flags);
 443         sg_cpu->last_update = time;
 444
 445         ignore_dl_rate_limit(sg_cpu, sg_policy);
 446
 447         if (!sugov_should_update_freq(sg_policy, time))
 448                 return;
 449
 450         busy = sugov_cpu_is_busy(sg_cpu);
 451
 452         util = sugov_get_util(sg_cpu);
 453         max = sg_cpu->max;
 454         util = sugov_iowait_apply(sg_cpu, time, util, max);
 455         next_f = get_next_freq(sg_policy, util, max);
 456         /*
 457          * Do not reduce the frequency if the CPU has not been idle
 458          * recently, as the reduction is likely to be premature then.
 459          */
 460         if (busy && next_f < sg_policy->next_freq) {
 461                 next_f = sg_policy->next_freq;
 462
 463                 /* Reset cached freq as next_freq has changed */
 464                 sg_policy->cached_raw_freq = 0;
 465         }
 466
 467         /*
 468          * This code runs under rq->lock for the target CPU, so it won't run
 469          * concurrently on two different CPUs for the same target and it is not
 470          * necessary to acquire the lock in the fast switch case.
 471          */
 472         if (sg_policy->policy->fast_switch_enabled) {
 473                 sugov_fast_switch(sg_policy, time, next_f);
 474         } else {
 475                 raw_spin_lock(&sg_policy->update_lock);
 476                 sugov_deferred_update(sg_policy, time, next_f);
 477                 raw_spin_unlock(&sg_policy->update_lock);
 478         }
 479 }
 480
 481 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
 482 {
 483         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 484         struct cpufreq_policy *policy = sg_policy->policy;
 485         unsigned long util = 0, max = 1;
 486         unsigned int j;
 487
 488         for_each_cpu(j, policy->cpus) {
 489                 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
 490                 unsigned long j_util, j_max;
 491
 492                 j_util = sugov_get_util(j_sg_cpu);
 493                 j_max = j_sg_cpu->max;
 494                 j_util = sugov_iowait_apply(j_sg_cpu, time, j_util, j_max);
 495
 496                 if (j_util * max > j_max * util) {
 497                         util = j_util;
 498                         max = j_max;
 499                 }
 500         }
 501
 502         return get_next_freq(sg_policy, util, max);
 503 }
 504
 505 static void
 506 sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
 507 {
 508         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 509         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 510         unsigned int next_f;
 511
 512         raw_spin_lock(&sg_policy->update_lock);
 513
 514         sugov_iowait_boost(sg_cpu, time, flags);
 515         sg_cpu->last_update = time;
 516
 517         ignore_dl_rate_limit(sg_cpu, sg_policy);
 518
 519         if (sugov_should_update_freq(sg_policy, time)) {
 520                 next_f = sugov_next_freq_shared(sg_cpu, time);
 521
 522                 if (sg_policy->policy->fast_switch_enabled)
 523                         sugov_fast_switch(sg_policy, time, next_f);
 524                 else
 525                         sugov_deferred_update(sg_policy, time, next_f);
 526         }
 527
 528         raw_spin_unlock(&sg_policy->update_lock);
 529 }
 530
 531 static void sugov_work(struct kthread_work *work)
 532 {
 533         struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
 534         unsigned int freq;
 535         unsigned long flags;
 536
 537         /*
 538          * Hold sg_policy->update_lock shortly to handle the case where:
 539          * incase sg_policy->next_freq is read here, and then updated by
 540          * sugov_deferred_update() just before work_in_progress is set to false
 541          * here, we may miss queueing the new update.
 542          *
 543          * Note: If a work was queued after the update_lock is released,
 544          * sugov_work() will just be called again by kthread_work code; and the
 545          * request will be proceed before the sugov thread sleeps.
 546          */
 547         raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
 548         freq = sg_policy->next_freq;
 549         sg_policy->work_in_progress = false;
 550         raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
 551
 552         mutex_lock(&sg_policy->work_lock);
 553         __cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
 554         mutex_unlock(&sg_policy->work_lock);
 555 }
 556
 557 static void sugov_irq_work(struct irq_work *irq_work)
 558 {
 559         struct sugov_policy *sg_policy;
 560
 561         sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
 562
 563         kthread_queue_work(&sg_policy->worker, &sg_policy->work);
 564 }
 565
 566 /************************** sysfs interface ************************/
 567
 568 static struct sugov_tunables *global_tunables;
 569 static DEFINE_MUTEX(global_tunables_lock);
 570
 571 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
 572 {
 573         return container_of(attr_set, struct sugov_tunables, attr_set);
 574 }
 575
 576 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
 577 {
 578         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 579
 580         return sprintf(buf, "%u\n", tunables->rate_limit_us);
 581 }
 582
 583 static ssize_t
 584 rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
 585 {
 586         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 587         struct sugov_policy *sg_policy;
 588         unsigned int rate_limit_us;
 589
 590         if (kstrtouint(buf, 10, &rate_limit_us))
 591                 return -EINVAL;
 592
 593         tunables->rate_limit_us = rate_limit_us;
 594
 595         list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
 596                 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
 597
 598         return count;
 599 }
 600
 601 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
 602
 603 static struct attribute *sugov_attrs[] = {
 604         &rate_limit_us.attr,
 605         NULL
 606 };
 607 ATTRIBUTE_GROUPS(sugov);
 608
 609 static struct kobj_type sugov_tunables_ktype = {
 610         .default_groups = sugov_groups,
 611         .sysfs_ops = &governor_sysfs_ops,
 612 };
 613
 614 /********************** cpufreq governor interface *********************/
 615
 616 struct cpufreq_governor schedutil_gov;
 617
 618 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
 619 {
 620         struct sugov_policy *sg_policy;
 621
 622         sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
 623         if (!sg_policy)
 624                 return NULL;
 625
 626         sg_policy->policy = policy;
 627         raw_spin_lock_init(&sg_policy->update_lock);
 628         return sg_policy;
 629 }
 630
 631 static void sugov_policy_free(struct sugov_policy *sg_policy)
 632 {
 633         kfree(sg_policy);
 634 }
 635
 636 static int sugov_kthread_create(struct sugov_policy *sg_policy)
 637 {
 638         struct task_struct *thread;
 639         struct sched_attr attr = {
 640                 .size           = sizeof(struct sched_attr),
 641                 .sched_policy   = SCHED_DEADLINE,
 642                 .sched_flags    = SCHED_FLAG_SUGOV,
 643                 .sched_nice     = 0,
 644                 .sched_priority = 0,
 645                 /*
 646                  * Fake (unused) bandwidth; workaround to "fix"
 647                  * priority inheritance.
 648                  */
 649                 .sched_runtime  =  1000000,
 650                 .sched_deadline = 10000000,
 651                 .sched_period   = 10000000,
 652         };
 653         struct cpufreq_policy *policy = sg_policy->policy;
 654         int ret;
 655
 656         /* kthread only required for slow path */
 657         if (policy->fast_switch_enabled)
 658                 return 0;
 659
 660         kthread_init_work(&sg_policy->work, sugov_work);
 661         kthread_init_worker(&sg_policy->worker);
 662         thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
 663                                 "sugov:%d",
 664                                 cpumask_first(policy->related_cpus));
 665         if (IS_ERR(thread)) {
 666                 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
 667                 return PTR_ERR(thread);
 668         }
 669
 670         ret = sched_setattr_nocheck(thread, &attr);
 671         if (ret) {
 672                 kthread_stop(thread);
 673                 pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
 674                 return ret;
 675         }
 676
 677         sg_policy->thread = thread;
 678         kthread_bind_mask(thread, policy->related_cpus);
 679         init_irq_work(&sg_policy->irq_work, sugov_irq_work);
 680         mutex_init(&sg_policy->work_lock);
 681
 682         wake_up_process(thread);
 683
 684         return 0;
 685 }
 686
 687 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
 688 {
 689         /* kthread only required for slow path */
 690         if (sg_policy->policy->fast_switch_enabled)
 691                 return;
 692
 693         kthread_flush_worker(&sg_policy->worker);
 694         kthread_stop(sg_policy->thread);
 695         mutex_destroy(&sg_policy->work_lock);
 696 }
 697
 698 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
 699 {
 700         struct sugov_tunables *tunables;
 701
 702         tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
 703         if (tunables) {
 704                 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
 705                 if (!have_governor_per_policy())
 706                         global_tunables = tunables;
 707         }
 708         return tunables;
 709 }
 710
 711 static void sugov_tunables_free(struct sugov_tunables *tunables)
 712 {
 713         if (!have_governor_per_policy())
 714                 global_tunables = NULL;
 715
 716         kfree(tunables);
 717 }
 718
 719 static int sugov_init(struct cpufreq_policy *policy)
 720 {
 721         struct sugov_policy *sg_policy;
 722         struct sugov_tunables *tunables;
 723         int ret = 0;
 724
 725         /* State should be equivalent to EXIT */
 726         if (policy->governor_data)
 727                 return -EBUSY;
 728
 729         cpufreq_enable_fast_switch(policy);
 730
 731         sg_policy = sugov_policy_alloc(policy);
 732         if (!sg_policy) {
 733                 ret = -ENOMEM;
 734                 goto disable_fast_switch;
 735         }
 736
 737         ret = sugov_kthread_create(sg_policy);
 738         if (ret)
 739                 goto free_sg_policy;
 740
 741         mutex_lock(&global_tunables_lock);
 742
 743         if (global_tunables) {
 744                 if (WARN_ON(have_governor_per_policy())) {
 745                         ret = -EINVAL;
 746                         goto stop_kthread;
 747                 }
 748                 policy->governor_data = sg_policy;
 749                 sg_policy->tunables = global_tunables;
 750
 751                 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
 752                 goto out;
 753         }
 754
 755         tunables = sugov_tunables_alloc(sg_policy);
 756         if (!tunables) {
 757                 ret = -ENOMEM;
 758                 goto stop_kthread;
 759         }
 760
 761         tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
 762
 763         policy->governor_data = sg_policy;
 764         sg_policy->tunables = tunables;
 765
 766         ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
 767                                    get_governor_parent_kobj(policy), "%s",
 768                                    schedutil_gov.name);
 769         if (ret)
 770                 goto fail;
 771
 772 out:
 773         mutex_unlock(&global_tunables_lock);
 774         return 0;
 775
 776 fail:
 777         kobject_put(&tunables->attr_set.kobj);
 778         policy->governor_data = NULL;
 779         sugov_tunables_free(tunables);
 780
 781 stop_kthread:
 782         sugov_kthread_stop(sg_policy);
 783         mutex_unlock(&global_tunables_lock);
 784
 785 free_sg_policy:
 786         sugov_policy_free(sg_policy);
 787
 788 disable_fast_switch:
 789         cpufreq_disable_fast_switch(policy);
 790
 791         pr_err("initialization failed (error %d)\n", ret);
 792         return ret;
 793 }
 794
 795 static void sugov_exit(struct cpufreq_policy *policy)
 796 {
 797         struct sugov_policy *sg_policy = policy->governor_data;
 798         struct sugov_tunables *tunables = sg_policy->tunables;
 799         unsigned int count;
 800
 801         mutex_lock(&global_tunables_lock);
 802
 803         count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
 804         policy->governor_data = NULL;
 805         if (!count)
 806                 sugov_tunables_free(tunables);
 807
 808         mutex_unlock(&global_tunables_lock);
 809
 810         sugov_kthread_stop(sg_policy);
 811         sugov_policy_free(sg_policy);
 812         cpufreq_disable_fast_switch(policy);
 813 }
 814
 815 static int sugov_start(struct cpufreq_policy *policy)
 816 {
 817         struct sugov_policy *sg_policy = policy->governor_data;
 818         unsigned int cpu;
 819
 820         sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
 821         sg_policy->last_freq_update_time        = 0;
 822         sg_policy->next_freq                    = 0;
 823         sg_policy->work_in_progress             = false;
 824         sg_policy->need_freq_update             = false;
 825         sg_policy->cached_raw_freq              = 0;
 826
 827         for_each_cpu(cpu, policy->cpus) {
 828                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 829
 830                 memset(sg_cpu, 0, sizeof(*sg_cpu));
 831                 sg_cpu->cpu                     = cpu;
 832                 sg_cpu->sg_policy               = sg_policy;
 833         }
 834
 835         for_each_cpu(cpu, policy->cpus) {
 836                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 837
 838                 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
 839                                              policy_is_shared(policy) ?
 840                                                         sugov_update_shared :
 841                                                         sugov_update_single);
 842         }
 843         return 0;
 844 }
 845
 846 static void sugov_stop(struct cpufreq_policy *policy)
 847 {
 848         struct sugov_policy *sg_policy = policy->governor_data;
 849         unsigned int cpu;
 850
 851         for_each_cpu(cpu, policy->cpus)
 852                 cpufreq_remove_update_util_hook(cpu);
 853
 854         synchronize_rcu();
 855
 856         if (!policy->fast_switch_enabled) {
 857                 irq_work_sync(&sg_policy->irq_work);
 858                 kthread_cancel_work_sync(&sg_policy->work);
 859         }
 860 }
 861
 862 static void sugov_limits(struct cpufreq_policy *policy)
 863 {
 864         struct sugov_policy *sg_policy = policy->governor_data;
 865
 866         if (!policy->fast_switch_enabled) {
 867                 mutex_lock(&sg_policy->work_lock);
 868                 cpufreq_policy_apply_limits(policy);
 869                 mutex_unlock(&sg_policy->work_lock);
 870         }
 871
 872         sg_policy->need_freq_update = true;
 873 }
 874
 875 struct cpufreq_governor schedutil_gov = {
 876         .name                   = "schedutil",
 877         .owner                  = THIS_MODULE,
 878         .dynamic_switching      = true,
 879         .init                   = sugov_init,
 880         .exit                   = sugov_exit,
 881         .start                  = sugov_start,
 882         .stop                   = sugov_stop,
 883         .limits                 = sugov_limits,
 884 };
 885
 886 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
 887 struct cpufreq_governor *cpufreq_default_governor(void)
 888 {
 889         return &schedutil_gov;
 890 }
 891 #endif
 892
 893 static int __init sugov_register(void)
 894 {
 895         return cpufreq_register_governor(&schedutil_gov);
 896 }
 897 fs_initcall(sugov_register);
 898
 899 #ifdef CONFIG_ENERGY_MODEL
 900 extern bool sched_energy_update;
 901 extern struct mutex sched_energy_mutex;
 902
 903 static void rebuild_sd_workfn(struct work_struct *work)
 904 {
 905         mutex_lock(&sched_energy_mutex);
 906         sched_energy_update = true;
 907         rebuild_sched_domains();
 908         sched_energy_update = false;
 909         mutex_unlock(&sched_energy_mutex);
 910 }
 911 static DECLARE_WORK(rebuild_sd_work, rebuild_sd_workfn);
 912
 913 /*
 914  * EAS shouldn't be attempted without sugov, so rebuild the sched_domains
 915  * on governor changes to make sure the scheduler knows about it.
 916  */
 917 void sched_cpufreq_governor_change(struct cpufreq_policy *policy,
 918                                   struct cpufreq_governor *old_gov)
 919 {
 920         if (old_gov == &schedutil_gov || policy->governor == &schedutil_gov) {
 921                 /*
 922                  * When called from the cpufreq_register_driver() path, the
 923                  * cpu_hotplug_lock is already held, so use a work item to
 924                  * avoid nested locking in rebuild_sched_domains().
 925                  */
 926                 schedule_work(&rebuild_sd_work);
 927         }
 928
 929 }
 930 #endif