kernel/sched/cpufreq_schedutil.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * CPUFreq governor based on scheduler-provided CPU utilization data.
   4  *
   5  * Copyright (C) 2016, Intel Corporation
   6  * Author: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
   7  */
   8
   9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  11 #include "sched.h"
  12
  13 #include <linux/sched/cpufreq.h>
  14 #include <trace/events/power.h>
  15
  16 #define IOWAIT_BOOST_MIN        (SCHED_CAPACITY_SCALE / 8)
  17
  18 struct sugov_tunables {
  19         struct gov_attr_set     attr_set;
  20         unsigned int            rate_limit_us;
  21 };
  22
  23 struct sugov_policy {
  24         struct cpufreq_policy   *policy;
  25
  26         struct sugov_tunables   *tunables;
  27         struct list_head        tunables_hook;
  28
  29         raw_spinlock_t          update_lock;    /* For shared policies */
  30         u64                     last_freq_update_time;
  31         s64                     freq_update_delay_ns;
  32         unsigned int            next_freq;
  33         unsigned int            cached_raw_freq;
  34
  35         /* The next fields are only needed if fast switch cannot be used: */
  36         struct                  irq_work irq_work;
  37         struct                  kthread_work work;
  38         struct                  mutex work_lock;
  39         struct                  kthread_worker worker;
  40         struct task_struct      *thread;
  41         bool                    work_in_progress;
  42
  43         bool                    limits_changed;
  44         bool                    need_freq_update;
  45 };
  46
  47 struct sugov_cpu {
  48         struct update_util_data update_util;
  49         struct sugov_policy     *sg_policy;
  50         unsigned int            cpu;
  51
  52         bool                    iowait_boost_pending;
  53         unsigned int            iowait_boost;
  54         u64                     last_update;
  55
  56         unsigned long           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          * For the slow switching platforms, the kthread is always scheduled on
  86          * the right set of CPUs and any CPU can find the next frequency and
  87          * schedule the kthread.
  88          */
  89         if (sg_policy->policy->fast_switch_enabled &&
  90             !cpufreq_this_cpu_can_update(sg_policy->policy))
  91                 return false;
  92
  93         if (unlikely(sg_policy->limits_changed)) {
  94                 sg_policy->limits_changed = false;
  95                 sg_policy->need_freq_update = true;
  96                 return true;
  97         }
  98
  99         delta_ns = time - sg_policy->last_freq_update_time;
 100
 101         return delta_ns >= sg_policy->freq_update_delay_ns;
 102 }
 103
 104 static bool sugov_update_next_freq(struct sugov_policy *sg_policy, u64 time,
 105                                    unsigned int next_freq)
 106 {
 107         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         struct cpufreq_policy *policy = sg_policy->policy;
 120
 121         if (!sugov_update_next_freq(sg_policy, time, next_freq))
 122                 return;
 123
 124         next_freq = cpufreq_driver_fast_switch(policy, next_freq);
 125         if (!next_freq)
 126                 return;
 127
 128         policy->cur = next_freq;
 129         trace_cpu_frequency(next_freq, smp_processor_id());
 130 }
 131
 132 static void sugov_deferred_update(struct sugov_policy *sg_policy, u64 time,
 133                                   unsigned int next_freq)
 134 {
 135         if (!sugov_update_next_freq(sg_policy, time, next_freq))
 136                 return;
 137
 138         if (!sg_policy->work_in_progress) {
 139                 sg_policy->work_in_progress = true;
 140                 irq_work_queue(&sg_policy->irq_work);
 141         }
 142 }
 143
 144 /**
 145  * get_next_freq - Compute a new frequency for a given cpufreq policy.
 146  * @sg_policy: schedutil policy object to compute the new frequency for.
 147  * @util: Current CPU utilization.
 148  * @max: CPU capacity.
 149  *
 150  * If the utilization is frequency-invariant, choose the new frequency to be
 151  * proportional to it, that is
 152  *
 153  * next_freq = C * max_freq * util / max
 154  *
 155  * Otherwise, approximate the would-be frequency-invariant utilization by
 156  * util_raw * (curr_freq / max_freq) which leads to
 157  *
 158  * next_freq = C * curr_freq * util_raw / max
 159  *
 160  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
 161  *
 162  * The lowest driver-supported frequency which is equal or greater than the raw
 163  * next_freq (as calculated above) is returned, subject to policy min/max and
 164  * cpufreq driver limitations.
 165  */
 166 static unsigned int get_next_freq(struct sugov_policy *sg_policy,
 167                                   unsigned long util, unsigned long max)
 168 {
 169         struct cpufreq_policy *policy = sg_policy->policy;
 170         unsigned int freq = arch_scale_freq_invariant() ?
 171                                 policy->cpuinfo.max_freq : policy->cur;
 172
 173         freq = map_util_freq(util, freq, max);
 174
 175         if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
 176                 return sg_policy->next_freq;
 177
 178         sg_policy->need_freq_update = false;
 179         sg_policy->cached_raw_freq = freq;
 180         return cpufreq_driver_resolve_freq(policy, freq);
 181 }
 182
 183 /*
 184  * This function computes an effective utilization for the given CPU, to be
 185  * used for frequency selection given the linear relation: f = u * f_max.
 186  *
 187  * The scheduler tracks the following metrics:
 188  *
 189  *   cpu_util_{cfs,rt,dl,irq}()
 190  *   cpu_bw_dl()
 191  *
 192  * Where the cfs,rt and dl util numbers are tracked with the same metric and
 193  * synchronized windows and are thus directly comparable.
 194  *
 195  * The cfs,rt,dl utilization are the running times measured with rq->clock_task
 196  * which excludes things like IRQ and steal-time. These latter are then accrued
 197  * in the irq utilization.
 198  *
 199  * The DL bandwidth number otoh is not a measured metric but a value computed
 200  * based on the task model parameters and gives the minimal utilization
 201  * required to meet deadlines.
 202  */
 203 unsigned long schedutil_cpu_util(int cpu, unsigned long util_cfs,
 204                                  unsigned long max, enum schedutil_type type,
 205                                  struct task_struct *p)
 206 {
 207         unsigned long dl_util, util, irq;
 208         struct rq *rq = cpu_rq(cpu);
 209
 210         if (!IS_BUILTIN(CONFIG_UCLAMP_TASK) &&
 211             type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
 212                 return max;
 213         }
 214
 215         /*
 216          * Early check to see if IRQ/steal time saturates the CPU, can be
 217          * because of inaccuracies in how we track these -- see
 218          * update_irq_load_avg().
 219          */
 220         irq = cpu_util_irq(rq);
 221         if (unlikely(irq >= max))
 222                 return max;
 223
 224         /*
 225          * Because the time spend on RT/DL tasks is visible as 'lost' time to
 226          * CFS tasks and we use the same metric to track the effective
 227          * utilization (PELT windows are synchronized) we can directly add them
 228          * to obtain the CPU's actual utilization.
 229          *
 230          * CFS and RT utilization can be boosted or capped, depending on
 231          * utilization clamp constraints requested by currently RUNNABLE
 232          * tasks.
 233          * When there are no CFS RUNNABLE tasks, clamps are released and
 234          * frequency will be gracefully reduced with the utilization decay.
 235          */
 236         util = util_cfs + cpu_util_rt(rq);
 237         if (type == FREQUENCY_UTIL)
 238                 util = uclamp_util_with(rq, util, p);
 239
 240         dl_util = cpu_util_dl(rq);
 241
 242         /*
 243          * For frequency selection we do not make cpu_util_dl() a permanent part
 244          * of this sum because we want to use cpu_bw_dl() later on, but we need
 245          * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
 246          * that we select f_max when there is no idle time.
 247          *
 248          * NOTE: numerical errors or stop class might cause us to not quite hit
 249          * saturation when we should -- something for later.
 250          */
 251         if (util + dl_util >= max)
 252                 return max;
 253
 254         /*
 255          * OTOH, for energy computation we need the estimated running time, so
 256          * include util_dl and ignore dl_bw.
 257          */
 258         if (type == ENERGY_UTIL)
 259                 util += dl_util;
 260
 261         /*
 262          * There is still idle time; further improve the number by using the
 263          * irq metric. Because IRQ/steal time is hidden from the task clock we
 264          * need to scale the task numbers:
 265          *
 266          *              max - irq
 267          *   U' = irq + --------- * U
 268          *                 max
 269          */
 270         util = scale_irq_capacity(util, irq, max);
 271         util += irq;
 272
 273         /*
 274          * Bandwidth required by DEADLINE must always be granted while, for
 275          * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
 276          * to gracefully reduce the frequency when no tasks show up for longer
 277          * periods of time.
 278          *
 279          * Ideally we would like to set bw_dl as min/guaranteed freq and util +
 280          * bw_dl as requested freq. However, cpufreq is not yet ready for such
 281          * an interface. So, we only do the latter for now.
 282          */
 283         if (type == FREQUENCY_UTIL)
 284                 util += cpu_bw_dl(rq);
 285
 286         return min(max, util);
 287 }
 288
 289 static unsigned long sugov_get_util(struct sugov_cpu *sg_cpu)
 290 {
 291         struct rq *rq = cpu_rq(sg_cpu->cpu);
 292         unsigned long util = cpu_util_cfs(rq);
 293         unsigned long max = arch_scale_cpu_capacity(sg_cpu->cpu);
 294
 295         sg_cpu->max = max;
 296         sg_cpu->bw_dl = cpu_bw_dl(rq);
 297
 298         return schedutil_cpu_util(sg_cpu->cpu, util, max, FREQUENCY_UTIL, NULL);
 299 }
 300
 301 /**
 302  * sugov_iowait_reset() - Reset the IO boost status of a CPU.
 303  * @sg_cpu: the sugov data for the CPU to boost
 304  * @time: the update time from the caller
 305  * @set_iowait_boost: true if an IO boost has been requested
 306  *
 307  * The IO wait boost of a task is disabled after a tick since the last update
 308  * of a CPU. If a new IO wait boost is requested after more then a tick, then
 309  * we enable the boost starting from IOWAIT_BOOST_MIN, which improves energy
 310  * efficiency by ignoring sporadic wakeups from IO.
 311  */
 312 static bool sugov_iowait_reset(struct sugov_cpu *sg_cpu, u64 time,
 313                                bool set_iowait_boost)
 314 {
 315         s64 delta_ns = time - sg_cpu->last_update;
 316
 317         /* Reset boost only if a tick has elapsed since last request */
 318         if (delta_ns <= TICK_NSEC)
 319                 return false;
 320
 321         sg_cpu->iowait_boost = set_iowait_boost ? IOWAIT_BOOST_MIN : 0;
 322         sg_cpu->iowait_boost_pending = set_iowait_boost;
 323
 324         return true;
 325 }
 326
 327 /**
 328  * sugov_iowait_boost() - Updates the IO boost status of a CPU.
 329  * @sg_cpu: the sugov data for the CPU to boost
 330  * @time: the update time from the caller
 331  * @flags: SCHED_CPUFREQ_IOWAIT if the task is waking up after an IO wait
 332  *
 333  * Each time a task wakes up after an IO operation, the CPU utilization can be
 334  * boosted to a certain utilization which doubles at each "frequent and
 335  * successive" wakeup from IO, ranging from IOWAIT_BOOST_MIN to the utilization
 336  * of the maximum OPP.
 337  *
 338  * To keep doubling, an IO boost has to be requested at least once per tick,
 339  * otherwise we restart from the utilization of the minimum OPP.
 340  */
 341 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
 342                                unsigned int flags)
 343 {
 344         bool set_iowait_boost = flags & SCHED_CPUFREQ_IOWAIT;
 345
 346         /* Reset boost if the CPU appears to have been idle enough */
 347         if (sg_cpu->iowait_boost &&
 348             sugov_iowait_reset(sg_cpu, time, set_iowait_boost))
 349                 return;
 350
 351         /* Boost only tasks waking up after IO */
 352         if (!set_iowait_boost)
 353                 return;
 354
 355         /* Ensure boost doubles only one time at each request */
 356         if (sg_cpu->iowait_boost_pending)
 357                 return;
 358         sg_cpu->iowait_boost_pending = true;
 359
 360         /* Double the boost at each request */
 361         if (sg_cpu->iowait_boost) {
 362                 sg_cpu->iowait_boost =
 363                         min_t(unsigned int, sg_cpu->iowait_boost << 1, SCHED_CAPACITY_SCALE);
 364                 return;
 365         }
 366
 367         /* First wakeup after IO: start with minimum boost */
 368         sg_cpu->iowait_boost = IOWAIT_BOOST_MIN;
 369 }
 370
 371 /**
 372  * sugov_iowait_apply() - Apply the IO boost to a CPU.
 373  * @sg_cpu: the sugov data for the cpu to boost
 374  * @time: the update time from the caller
 375  * @util: the utilization to (eventually) boost
 376  * @max: the maximum value the utilization can be boosted to
 377  *
 378  * A CPU running a task which woken up after an IO operation can have its
 379  * utilization boosted to speed up the completion of those IO operations.
 380  * The IO boost value is increased each time a task wakes up from IO, in
 381  * sugov_iowait_apply(), and it's instead decreased by this function,
 382  * each time an increase has not been requested (!iowait_boost_pending).
 383  *
 384  * A CPU which also appears to have been idle for at least one tick has also
 385  * its IO boost utilization reset.
 386  *
 387  * This mechanism is designed to boost high frequently IO waiting tasks, while
 388  * being more conservative on tasks which does sporadic IO operations.
 389  */
 390 static unsigned long sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
 391                                         unsigned long util, unsigned long max)
 392 {
 393         unsigned long boost;
 394
 395         /* No boost currently required */
 396         if (!sg_cpu->iowait_boost)
 397                 return util;
 398
 399         /* Reset boost if the CPU appears to have been idle enough */
 400         if (sugov_iowait_reset(sg_cpu, time, false))
 401                 return util;
 402
 403         if (!sg_cpu->iowait_boost_pending) {
 404                 /*
 405                  * No boost pending; reduce the boost value.
 406                  */
 407                 sg_cpu->iowait_boost >>= 1;
 408                 if (sg_cpu->iowait_boost < IOWAIT_BOOST_MIN) {
 409                         sg_cpu->iowait_boost = 0;
 410                         return util;
 411                 }
 412         }
 413
 414         sg_cpu->iowait_boost_pending = false;
 415
 416         /*
 417          * @util is already in capacity scale; convert iowait_boost
 418          * into the same scale so we can compare.
 419          */
 420         boost = (sg_cpu->iowait_boost * max) >> SCHED_CAPACITY_SHIFT;
 421         return max(boost, util);
 422 }
 423
 424 #ifdef CONFIG_NO_HZ_COMMON
 425 static bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu)
 426 {
 427         unsigned long idle_calls = tick_nohz_get_idle_calls_cpu(sg_cpu->cpu);
 428         bool ret = idle_calls == sg_cpu->saved_idle_calls;
 429
 430         sg_cpu->saved_idle_calls = idle_calls;
 431         return ret;
 432 }
 433 #else
 434 static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
 435 #endif /* CONFIG_NO_HZ_COMMON */
 436
 437 /*
 438  * Make sugov_should_update_freq() ignore the rate limit when DL
 439  * has increased the utilization.
 440  */
 441 static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu, struct sugov_policy *sg_policy)
 442 {
 443         if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_dl)
 444                 sg_policy->limits_changed = true;
 445 }
 446
 447 static void sugov_update_single(struct update_util_data *hook, u64 time,
 448                                 unsigned int flags)
 449 {
 450         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 451         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 452         unsigned long util, max;
 453         unsigned int next_f;
 454         bool busy;
 455
 456         sugov_iowait_boost(sg_cpu, time, flags);
 457         sg_cpu->last_update = time;
 458
 459         ignore_dl_rate_limit(sg_cpu, sg_policy);
 460
 461         if (!sugov_should_update_freq(sg_policy, time))
 462                 return;
 463
 464         /* Limits may have changed, don't skip frequency update */
 465         busy = !sg_policy->need_freq_update && sugov_cpu_is_busy(sg_cpu);
 466
 467         util = sugov_get_util(sg_cpu);
 468         max = sg_cpu->max;
 469         util = sugov_iowait_apply(sg_cpu, time, util, max);
 470         next_f = get_next_freq(sg_policy, util, max);
 471         /*
 472          * Do not reduce the frequency if the CPU has not been idle
 473          * recently, as the reduction is likely to be premature then.
 474          */
 475         if (busy && next_f < sg_policy->next_freq) {
 476                 next_f = sg_policy->next_freq;
 477
 478                 /* Reset cached freq as next_freq has changed */
 479                 sg_policy->cached_raw_freq = 0;
 480         }
 481
 482         /*
 483          * This code runs under rq->lock for the target CPU, so it won't run
 484          * concurrently on two different CPUs for the same target and it is not
 485          * necessary to acquire the lock in the fast switch case.
 486          */
 487         if (sg_policy->policy->fast_switch_enabled) {
 488                 sugov_fast_switch(sg_policy, time, next_f);
 489         } else {
 490                 raw_spin_lock(&sg_policy->update_lock);
 491                 sugov_deferred_update(sg_policy, time, next_f);
 492                 raw_spin_unlock(&sg_policy->update_lock);
 493         }
 494 }
 495
 496 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
 497 {
 498         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 499         struct cpufreq_policy *policy = sg_policy->policy;
 500         unsigned long util = 0, max = 1;
 501         unsigned int j;
 502
 503         for_each_cpu(j, policy->cpus) {
 504                 struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
 505                 unsigned long j_util, j_max;
 506
 507                 j_util = sugov_get_util(j_sg_cpu);
 508                 j_max = j_sg_cpu->max;
 509                 j_util = sugov_iowait_apply(j_sg_cpu, time, j_util, j_max);
 510
 511                 if (j_util * max > j_max * util) {
 512                         util = j_util;
 513                         max = j_max;
 514                 }
 515         }
 516
 517         return get_next_freq(sg_policy, util, max);
 518 }
 519
 520 static void
 521 sugov_update_shared(struct update_util_data *hook, u64 time, unsigned int flags)
 522 {
 523         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 524         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 525         unsigned int next_f;
 526
 527         raw_spin_lock(&sg_policy->update_lock);
 528
 529         sugov_iowait_boost(sg_cpu, time, flags);
 530         sg_cpu->last_update = time;
 531
 532         ignore_dl_rate_limit(sg_cpu, sg_policy);
 533
 534         if (sugov_should_update_freq(sg_policy, time)) {
 535                 next_f = sugov_next_freq_shared(sg_cpu, time);
 536
 537                 if (sg_policy->policy->fast_switch_enabled)
 538                         sugov_fast_switch(sg_policy, time, next_f);
 539                 else
 540                         sugov_deferred_update(sg_policy, time, next_f);
 541         }
 542
 543         raw_spin_unlock(&sg_policy->update_lock);
 544 }
 545
 546 static void sugov_work(struct kthread_work *work)
 547 {
 548         struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
 549         unsigned int freq;
 550         unsigned long flags;
 551
 552         /*
 553          * Hold sg_policy->update_lock shortly to handle the case where:
 554          * incase sg_policy->next_freq is read here, and then updated by
 555          * sugov_deferred_update() just before work_in_progress is set to false
 556          * here, we may miss queueing the new update.
 557          *
 558          * Note: If a work was queued after the update_lock is released,
 559          * sugov_work() will just be called again by kthread_work code; and the
 560          * request will be proceed before the sugov thread sleeps.
 561          */
 562         raw_spin_lock_irqsave(&sg_policy->update_lock, flags);
 563         freq = sg_policy->next_freq;
 564         sg_policy->work_in_progress = false;
 565         raw_spin_unlock_irqrestore(&sg_policy->update_lock, flags);
 566
 567         mutex_lock(&sg_policy->work_lock);
 568         __cpufreq_driver_target(sg_policy->policy, freq, CPUFREQ_RELATION_L);
 569         mutex_unlock(&sg_policy->work_lock);
 570 }
 571
 572 static void sugov_irq_work(struct irq_work *irq_work)
 573 {
 574         struct sugov_policy *sg_policy;
 575
 576         sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
 577
 578         kthread_queue_work(&sg_policy->worker, &sg_policy->work);
 579 }
 580
 581 /************************** sysfs interface ************************/
 582
 583 static struct sugov_tunables *global_tunables;
 584 static DEFINE_MUTEX(global_tunables_lock);
 585
 586 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
 587 {
 588         return container_of(attr_set, struct sugov_tunables, attr_set);
 589 }
 590
 591 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
 592 {
 593         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 594
 595         return sprintf(buf, "%u\n", tunables->rate_limit_us);
 596 }
 597
 598 static ssize_t
 599 rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf, size_t count)
 600 {
 601         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 602         struct sugov_policy *sg_policy;
 603         unsigned int rate_limit_us;
 604
 605         if (kstrtouint(buf, 10, &rate_limit_us))
 606                 return -EINVAL;
 607
 608         tunables->rate_limit_us = rate_limit_us;
 609
 610         list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
 611                 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
 612
 613         return count;
 614 }
 615
 616 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
 617
 618 static struct attribute *sugov_attrs[] = {
 619         &rate_limit_us.attr,
 620         NULL
 621 };
 622 ATTRIBUTE_GROUPS(sugov);
 623
 624 static struct kobj_type sugov_tunables_ktype = {
 625         .default_groups = sugov_groups,
 626         .sysfs_ops = &governor_sysfs_ops,
 627 };
 628
 629 /********************** cpufreq governor interface *********************/
 630
 631 struct cpufreq_governor schedutil_gov;
 632
 633 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
 634 {
 635         struct sugov_policy *sg_policy;
 636
 637         sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
 638         if (!sg_policy)
 639                 return NULL;
 640
 641         sg_policy->policy = policy;
 642         raw_spin_lock_init(&sg_policy->update_lock);
 643         return sg_policy;
 644 }
 645
 646 static void sugov_policy_free(struct sugov_policy *sg_policy)
 647 {
 648         kfree(sg_policy);
 649 }
 650
 651 static int sugov_kthread_create(struct sugov_policy *sg_policy)
 652 {
 653         struct task_struct *thread;
 654         struct sched_attr attr = {
 655                 .size           = sizeof(struct sched_attr),
 656                 .sched_policy   = SCHED_DEADLINE,
 657                 .sched_flags    = SCHED_FLAG_SUGOV,
 658                 .sched_nice     = 0,
 659                 .sched_priority = 0,
 660                 /*
 661                  * Fake (unused) bandwidth; workaround to "fix"
 662                  * priority inheritance.
 663                  */
 664                 .sched_runtime  =  1000000,
 665                 .sched_deadline = 10000000,
 666                 .sched_period   = 10000000,
 667         };
 668         struct cpufreq_policy *policy = sg_policy->policy;
 669         int ret;
 670
 671         /* kthread only required for slow path */
 672         if (policy->fast_switch_enabled)
 673                 return 0;
 674
 675         kthread_init_work(&sg_policy->work, sugov_work);
 676         kthread_init_worker(&sg_policy->worker);
 677         thread = kthread_create(kthread_worker_fn, &sg_policy->worker,
 678                                 "sugov:%d",
 679                                 cpumask_first(policy->related_cpus));
 680         if (IS_ERR(thread)) {
 681                 pr_err("failed to create sugov thread: %ld\n", PTR_ERR(thread));
 682                 return PTR_ERR(thread);
 683         }
 684
 685         ret = sched_setattr_nocheck(thread, &attr);
 686         if (ret) {
 687                 kthread_stop(thread);
 688                 pr_warn("%s: failed to set SCHED_DEADLINE\n", __func__);
 689                 return ret;
 690         }
 691
 692         sg_policy->thread = thread;
 693         kthread_bind_mask(thread, policy->related_cpus);
 694         init_irq_work(&sg_policy->irq_work, sugov_irq_work);
 695         mutex_init(&sg_policy->work_lock);
 696
 697         wake_up_process(thread);
 698
 699         return 0;
 700 }
 701
 702 static void sugov_kthread_stop(struct sugov_policy *sg_policy)
 703 {
 704         /* kthread only required for slow path */
 705         if (sg_policy->policy->fast_switch_enabled)
 706                 return;
 707
 708         kthread_flush_worker(&sg_policy->worker);
 709         kthread_stop(sg_policy->thread);
 710         mutex_destroy(&sg_policy->work_lock);
 711 }
 712
 713 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
 714 {
 715         struct sugov_tunables *tunables;
 716
 717         tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
 718         if (tunables) {
 719                 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
 720                 if (!have_governor_per_policy())
 721                         global_tunables = tunables;
 722         }
 723         return tunables;
 724 }
 725
 726 static void sugov_tunables_free(struct sugov_tunables *tunables)
 727 {
 728         if (!have_governor_per_policy())
 729                 global_tunables = NULL;
 730
 731         kfree(tunables);
 732 }
 733
 734 static int sugov_init(struct cpufreq_policy *policy)
 735 {
 736         struct sugov_policy *sg_policy;
 737         struct sugov_tunables *tunables;
 738         int ret = 0;
 739
 740         /* State should be equivalent to EXIT */
 741         if (policy->governor_data)
 742                 return -EBUSY;
 743
 744         cpufreq_enable_fast_switch(policy);
 745
 746         sg_policy = sugov_policy_alloc(policy);
 747         if (!sg_policy) {
 748                 ret = -ENOMEM;
 749                 goto disable_fast_switch;
 750         }
 751
 752         ret = sugov_kthread_create(sg_policy);
 753         if (ret)
 754                 goto free_sg_policy;
 755
 756         mutex_lock(&global_tunables_lock);
 757
 758         if (global_tunables) {
 759                 if (WARN_ON(have_governor_per_policy())) {
 760                         ret = -EINVAL;
 761                         goto stop_kthread;
 762                 }
 763                 policy->governor_data = sg_policy;
 764                 sg_policy->tunables = global_tunables;
 765
 766                 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
 767                 goto out;
 768         }
 769
 770         tunables = sugov_tunables_alloc(sg_policy);
 771         if (!tunables) {
 772                 ret = -ENOMEM;
 773                 goto stop_kthread;
 774         }
 775
 776         tunables->rate_limit_us = cpufreq_policy_transition_delay_us(policy);
 777
 778         policy->governor_data = sg_policy;
 779         sg_policy->tunables = tunables;
 780
 781         ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
 782                                    get_governor_parent_kobj(policy), "%s",
 783                                    schedutil_gov.name);
 784         if (ret)
 785                 goto fail;
 786
 787 out:
 788         mutex_unlock(&global_tunables_lock);
 789         return 0;
 790
 791 fail:
 792         kobject_put(&tunables->attr_set.kobj);
 793         policy->governor_data = NULL;
 794         sugov_tunables_free(tunables);
 795
 796 stop_kthread:
 797         sugov_kthread_stop(sg_policy);
 798         mutex_unlock(&global_tunables_lock);
 799
 800 free_sg_policy:
 801         sugov_policy_free(sg_policy);
 802
 803 disable_fast_switch:
 804         cpufreq_disable_fast_switch(policy);
 805
 806         pr_err("initialization failed (error %d)\n", ret);
 807         return ret;
 808 }
 809
 810 static void sugov_exit(struct cpufreq_policy *policy)
 811 {
 812         struct sugov_policy *sg_policy = policy->governor_data;
 813         struct sugov_tunables *tunables = sg_policy->tunables;
 814         unsigned int count;
 815
 816         mutex_lock(&global_tunables_lock);
 817
 818         count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
 819         policy->governor_data = NULL;
 820         if (!count)
 821                 sugov_tunables_free(tunables);
 822
 823         mutex_unlock(&global_tunables_lock);
 824
 825         sugov_kthread_stop(sg_policy);
 826         sugov_policy_free(sg_policy);
 827         cpufreq_disable_fast_switch(policy);
 828 }
 829
 830 static int sugov_start(struct cpufreq_policy *policy)
 831 {
 832         struct sugov_policy *sg_policy = policy->governor_data;
 833         unsigned int cpu;
 834
 835         sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
 836         sg_policy->last_freq_update_time        = 0;
 837         sg_policy->next_freq                    = 0;
 838         sg_policy->work_in_progress             = false;
 839         sg_policy->limits_changed               = false;
 840         sg_policy->need_freq_update             = false;
 841         sg_policy->cached_raw_freq              = 0;
 842
 843         for_each_cpu(cpu, policy->cpus) {
 844                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 845
 846                 memset(sg_cpu, 0, sizeof(*sg_cpu));
 847                 sg_cpu->cpu                     = cpu;
 848                 sg_cpu->sg_policy               = sg_policy;
 849         }
 850
 851         for_each_cpu(cpu, policy->cpus) {
 852                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 853
 854                 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
 855                                              policy_is_shared(policy) ?
 856                                                         sugov_update_shared :
 857                                                         sugov_update_single);
 858         }
 859         return 0;
 860 }
 861
 862 static void sugov_stop(struct cpufreq_policy *policy)
 863 {
 864         struct sugov_policy *sg_policy = policy->governor_data;
 865         unsigned int cpu;
 866
 867         for_each_cpu(cpu, policy->cpus)
 868                 cpufreq_remove_update_util_hook(cpu);
 869
 870         synchronize_rcu();
 871
 872         if (!policy->fast_switch_enabled) {
 873                 irq_work_sync(&sg_policy->irq_work);
 874                 kthread_cancel_work_sync(&sg_policy->work);
 875         }
 876 }
 877
 878 static void sugov_limits(struct cpufreq_policy *policy)
 879 {
 880         struct sugov_policy *sg_policy = policy->governor_data;
 881
 882         if (!policy->fast_switch_enabled) {
 883                 mutex_lock(&sg_policy->work_lock);
 884                 cpufreq_policy_apply_limits(policy);
 885                 mutex_unlock(&sg_policy->work_lock);
 886         }
 887
 888         sg_policy->limits_changed = true;
 889 }
 890
 891 struct cpufreq_governor schedutil_gov = {
 892         .name                   = "schedutil",
 893         .owner                  = THIS_MODULE,
 894         .dynamic_switching      = true,
 895         .init                   = sugov_init,
 896         .exit                   = sugov_exit,
 897         .start                  = sugov_start,
 898         .stop                   = sugov_stop,
 899         .limits                 = sugov_limits,
 900 };
 901
 902 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
 903 struct cpufreq_governor *cpufreq_default_governor(void)
 904 {
 905         return &schedutil_gov;
 906 }
 907 #endif
 908
 909 static int __init sugov_register(void)
 910 {
 911         return cpufreq_register_governor(&schedutil_gov);
 912 }
 913 fs_initcall(sugov_register);
 914
 915 #ifdef CONFIG_ENERGY_MODEL
 916 extern bool sched_energy_update;
 917 extern struct mutex sched_energy_mutex;
 918
 919 static void rebuild_sd_workfn(struct work_struct *work)
 920 {
 921         mutex_lock(&sched_energy_mutex);
 922         sched_energy_update = true;
 923         rebuild_sched_domains();
 924         sched_energy_update = false;
 925         mutex_unlock(&sched_energy_mutex);
 926 }
 927 static DECLARE_WORK(rebuild_sd_work, rebuild_sd_workfn);
 928
 929 /*
 930  * EAS shouldn't be attempted without sugov, so rebuild the sched_domains
 931  * on governor changes to make sure the scheduler knows about it.
 932  */
 933 void sched_cpufreq_governor_change(struct cpufreq_policy *policy,
 934                                   struct cpufreq_governor *old_gov)
 935 {
 936         if (old_gov == &schedutil_gov || policy->governor == &schedutil_gov) {
 937                 /*
 938                  * When called from the cpufreq_register_driver() path, the
 939                  * cpu_hotplug_lock is already held, so use a work item to
 940                  * avoid nested locking in rebuild_sched_domains().
 941                  */
 942                 schedule_work(&rebuild_sd_work);
 943         }
 944
 945 }
 946 #endif