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