kernel/sched/cpufreq_schedutil.c

   1 /*
   2  * CPUFreq governor based on scheduler-provided CPU utilization data.
   3  *
   4  * Copyright (C) 2016, Intel Corporation
   5  * Author: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
   6  *
   7  * This program is free software; you can redistribute it and/or modify
   8  * it under the terms of the GNU General Public License version 2 as
   9  * published by the Free Software Foundation.
  10  */
  11
  12 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  13
  14 #include <linux/cpufreq.h>
  15 #include <linux/slab.h>
  16 #include <trace/events/power.h>
  17
  18 #include "sched.h"
  19
  20 struct sugov_tunables {
  21         struct gov_attr_set attr_set;
  22         unsigned int rate_limit_us;
  23 };
  24
  25 struct sugov_policy {
  26         struct cpufreq_policy *policy;
  27
  28         struct sugov_tunables *tunables;
  29         struct list_head tunables_hook;
  30
  31         raw_spinlock_t update_lock;  /* For shared policies */
  32         u64 last_freq_update_time;
  33         s64 freq_update_delay_ns;
  34         unsigned int next_freq;
  35         unsigned int cached_raw_freq;
  36
  37         /* The next fields are only needed if fast switch cannot be used. */
  38         struct irq_work irq_work;
  39         struct work_struct work;
  40         struct mutex work_lock;
  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
  50         unsigned long iowait_boost;
  51         unsigned long iowait_boost_max;
  52         u64 last_update;
  53
  54         /* The fields below are only needed when sharing a policy. */
  55         unsigned long util;
  56         unsigned long max;
  57         unsigned int flags;
  58 };
  59
  60 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
  61
  62 /************************ Governor internals ***********************/
  63
  64 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
  65 {
  66         s64 delta_ns;
  67
  68         if (sg_policy->work_in_progress)
  69                 return false;
  70
  71         if (unlikely(sg_policy->need_freq_update)) {
  72                 sg_policy->need_freq_update = false;
  73                 /*
  74                  * This happens when limits change, so forget the previous
  75                  * next_freq value and force an update.
  76                  */
  77                 sg_policy->next_freq = UINT_MAX;
  78                 return true;
  79         }
  80
  81         delta_ns = time - sg_policy->last_freq_update_time;
  82         return delta_ns >= sg_policy->freq_update_delay_ns;
  83 }
  84
  85 static void sugov_update_commit(struct sugov_policy *sg_policy, u64 time,
  86                                 unsigned int next_freq)
  87 {
  88         struct cpufreq_policy *policy = sg_policy->policy;
  89
  90         sg_policy->last_freq_update_time = time;
  91
  92         if (policy->fast_switch_enabled) {
  93                 if (sg_policy->next_freq == next_freq) {
  94                         trace_cpu_frequency(policy->cur, smp_processor_id());
  95                         return;
  96                 }
  97                 sg_policy->next_freq = next_freq;
  98                 next_freq = cpufreq_driver_fast_switch(policy, next_freq);
  99                 if (next_freq == CPUFREQ_ENTRY_INVALID)
 100                         return;
 101
 102                 policy->cur = next_freq;
 103                 trace_cpu_frequency(next_freq, smp_processor_id());
 104         } else if (sg_policy->next_freq != next_freq) {
 105                 sg_policy->next_freq = next_freq;
 106                 sg_policy->work_in_progress = true;
 107                 irq_work_queue(&sg_policy->irq_work);
 108         }
 109 }
 110
 111 /**
 112  * get_next_freq - Compute a new frequency for a given cpufreq policy.
 113  * @sg_cpu: schedutil cpu object to compute the new frequency for.
 114  * @util: Current CPU utilization.
 115  * @max: CPU capacity.
 116  *
 117  * If the utilization is frequency-invariant, choose the new frequency to be
 118  * proportional to it, that is
 119  *
 120  * next_freq = C * max_freq * util / max
 121  *
 122  * Otherwise, approximate the would-be frequency-invariant utilization by
 123  * util_raw * (curr_freq / max_freq) which leads to
 124  *
 125  * next_freq = C * curr_freq * util_raw / max
 126  *
 127  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
 128  *
 129  * The lowest driver-supported frequency which is equal or greater than the raw
 130  * next_freq (as calculated above) is returned, subject to policy min/max and
 131  * cpufreq driver limitations.
 132  */
 133 static unsigned int get_next_freq(struct sugov_cpu *sg_cpu, unsigned long util,
 134                                   unsigned long max)
 135 {
 136         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 137         struct cpufreq_policy *policy = sg_policy->policy;
 138         unsigned int freq = arch_scale_freq_invariant() ?
 139                                 policy->cpuinfo.max_freq : policy->cur;
 140
 141         freq = (freq + (freq >> 2)) * util / max;
 142
 143         if (freq == sg_policy->cached_raw_freq && sg_policy->next_freq != UINT_MAX)
 144                 return sg_policy->next_freq;
 145         sg_policy->cached_raw_freq = freq;
 146         return cpufreq_driver_resolve_freq(policy, freq);
 147 }
 148
 149 static void sugov_get_util(unsigned long *util, unsigned long *max)
 150 {
 151         struct rq *rq = this_rq();
 152         unsigned long cfs_max;
 153
 154         cfs_max = arch_scale_cpu_capacity(NULL, smp_processor_id());
 155
 156         *util = min(rq->cfs.avg.util_avg, cfs_max);
 157         *max = cfs_max;
 158 }
 159
 160 static void sugov_set_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
 161                                    unsigned int flags)
 162 {
 163         if (flags & SCHED_CPUFREQ_IOWAIT) {
 164                 sg_cpu->iowait_boost = sg_cpu->iowait_boost_max;
 165         } else if (sg_cpu->iowait_boost) {
 166                 s64 delta_ns = time - sg_cpu->last_update;
 167
 168                 /* Clear iowait_boost if the CPU apprears to have been idle. */
 169                 if (delta_ns > TICK_NSEC)
 170                         sg_cpu->iowait_boost = 0;
 171         }
 172 }
 173
 174 static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, unsigned long *util,
 175                                unsigned long *max)
 176 {
 177         unsigned long boost_util = sg_cpu->iowait_boost;
 178         unsigned long boost_max = sg_cpu->iowait_boost_max;
 179
 180         if (!boost_util)
 181                 return;
 182
 183         if (*util * boost_max < *max * boost_util) {
 184                 *util = boost_util;
 185                 *max = boost_max;
 186         }
 187         sg_cpu->iowait_boost >>= 1;
 188 }
 189
 190 static void sugov_update_single(struct update_util_data *hook, u64 time,
 191                                 unsigned int flags)
 192 {
 193         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 194         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 195         struct cpufreq_policy *policy = sg_policy->policy;
 196         unsigned long util, max;
 197         unsigned int next_f;
 198
 199         sugov_set_iowait_boost(sg_cpu, time, flags);
 200         sg_cpu->last_update = time;
 201
 202         if (!sugov_should_update_freq(sg_policy, time))
 203                 return;
 204
 205         if (flags & SCHED_CPUFREQ_RT_DL) {
 206                 next_f = policy->cpuinfo.max_freq;
 207         } else {
 208                 sugov_get_util(&util, &max);
 209                 sugov_iowait_boost(sg_cpu, &util, &max);
 210                 next_f = get_next_freq(sg_cpu, util, max);
 211         }
 212         sugov_update_commit(sg_policy, time, next_f);
 213 }
 214
 215 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu,
 216                                            unsigned long util, unsigned long max,
 217                                            unsigned int flags)
 218 {
 219         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 220         struct cpufreq_policy *policy = sg_policy->policy;
 221         unsigned int max_f = policy->cpuinfo.max_freq;
 222         u64 last_freq_update_time = sg_policy->last_freq_update_time;
 223         unsigned int j;
 224
 225         if (flags & SCHED_CPUFREQ_RT_DL)
 226                 return max_f;
 227
 228         sugov_iowait_boost(sg_cpu, &util, &max);
 229
 230         for_each_cpu(j, policy->cpus) {
 231                 struct sugov_cpu *j_sg_cpu;
 232                 unsigned long j_util, j_max;
 233                 s64 delta_ns;
 234
 235                 if (j == smp_processor_id())
 236                         continue;
 237
 238                 j_sg_cpu = &per_cpu(sugov_cpu, j);
 239                 /*
 240                  * If the CPU utilization was last updated before the previous
 241                  * frequency update and the time elapsed between the last update
 242                  * of the CPU utilization and the last frequency update is long
 243                  * enough, don't take the CPU into account as it probably is
 244                  * idle now (and clear iowait_boost for it).
 245                  */
 246                 delta_ns = last_freq_update_time - j_sg_cpu->last_update;
 247                 if (delta_ns > TICK_NSEC) {
 248                         j_sg_cpu->iowait_boost = 0;
 249                         continue;
 250                 }
 251                 if (j_sg_cpu->flags & SCHED_CPUFREQ_RT_DL)
 252                         return max_f;
 253
 254                 j_util = j_sg_cpu->util;
 255                 j_max = j_sg_cpu->max;
 256                 if (j_util * max > j_max * util) {
 257                         util = j_util;
 258                         max = j_max;
 259                 }
 260
 261                 sugov_iowait_boost(j_sg_cpu, &util, &max);
 262         }
 263
 264         return get_next_freq(sg_cpu, util, max);
 265 }
 266
 267 static void sugov_update_shared(struct update_util_data *hook, u64 time,
 268                                 unsigned int flags)
 269 {
 270         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 271         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 272         unsigned long util, max;
 273         unsigned int next_f;
 274
 275         sugov_get_util(&util, &max);
 276
 277         raw_spin_lock(&sg_policy->update_lock);
 278
 279         sg_cpu->util = util;
 280         sg_cpu->max = max;
 281         sg_cpu->flags = flags;
 282
 283         sugov_set_iowait_boost(sg_cpu, time, flags);
 284         sg_cpu->last_update = time;
 285
 286         if (sugov_should_update_freq(sg_policy, time)) {
 287                 next_f = sugov_next_freq_shared(sg_cpu, util, max, flags);
 288                 sugov_update_commit(sg_policy, time, next_f);
 289         }
 290
 291         raw_spin_unlock(&sg_policy->update_lock);
 292 }
 293
 294 static void sugov_work(struct work_struct *work)
 295 {
 296         struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
 297
 298         mutex_lock(&sg_policy->work_lock);
 299         __cpufreq_driver_target(sg_policy->policy, sg_policy->next_freq,
 300                                 CPUFREQ_RELATION_L);
 301         mutex_unlock(&sg_policy->work_lock);
 302
 303         sg_policy->work_in_progress = false;
 304 }
 305
 306 static void sugov_irq_work(struct irq_work *irq_work)
 307 {
 308         struct sugov_policy *sg_policy;
 309
 310         sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
 311         schedule_work_on(smp_processor_id(), &sg_policy->work);
 312 }
 313
 314 /************************** sysfs interface ************************/
 315
 316 static struct sugov_tunables *global_tunables;
 317 static DEFINE_MUTEX(global_tunables_lock);
 318
 319 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
 320 {
 321         return container_of(attr_set, struct sugov_tunables, attr_set);
 322 }
 323
 324 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
 325 {
 326         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 327
 328         return sprintf(buf, "%u\n", tunables->rate_limit_us);
 329 }
 330
 331 static ssize_t rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf,
 332                                    size_t count)
 333 {
 334         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 335         struct sugov_policy *sg_policy;
 336         unsigned int rate_limit_us;
 337
 338         if (kstrtouint(buf, 10, &rate_limit_us))
 339                 return -EINVAL;
 340
 341         tunables->rate_limit_us = rate_limit_us;
 342
 343         list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
 344                 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
 345
 346         return count;
 347 }
 348
 349 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
 350
 351 static struct attribute *sugov_attributes[] = {
 352         &rate_limit_us.attr,
 353         NULL
 354 };
 355
 356 static void sugov_tunables_free(struct kobject *kobj)
 357 {
 358         struct gov_attr_set *attr_set = container_of(kobj, struct gov_attr_set, kobj);
 359
 360         kfree(to_sugov_tunables(attr_set));
 361 }
 362
 363 static struct kobj_type sugov_tunables_ktype = {
 364         .default_attrs = sugov_attributes,
 365         .sysfs_ops = &governor_sysfs_ops,
 366         .release = &sugov_tunables_free,
 367 };
 368
 369 /********************** cpufreq governor interface *********************/
 370
 371 static struct cpufreq_governor schedutil_gov;
 372
 373 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
 374 {
 375         struct sugov_policy *sg_policy;
 376
 377         sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
 378         if (!sg_policy)
 379                 return NULL;
 380
 381         sg_policy->policy = policy;
 382         init_irq_work(&sg_policy->irq_work, sugov_irq_work);
 383         INIT_WORK(&sg_policy->work, sugov_work);
 384         mutex_init(&sg_policy->work_lock);
 385         raw_spin_lock_init(&sg_policy->update_lock);
 386         return sg_policy;
 387 }
 388
 389 static void sugov_policy_free(struct sugov_policy *sg_policy)
 390 {
 391         mutex_destroy(&sg_policy->work_lock);
 392         kfree(sg_policy);
 393 }
 394
 395 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
 396 {
 397         struct sugov_tunables *tunables;
 398
 399         tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
 400         if (tunables) {
 401                 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
 402                 if (!have_governor_per_policy())
 403                         global_tunables = tunables;
 404         }
 405         return tunables;
 406 }
 407
 408 static void sugov_clear_global_tunables(void)
 409 {
 410         if (!have_governor_per_policy())
 411                 global_tunables = NULL;
 412 }
 413
 414 static int sugov_init(struct cpufreq_policy *policy)
 415 {
 416         struct sugov_policy *sg_policy;
 417         struct sugov_tunables *tunables;
 418         unsigned int lat;
 419         int ret = 0;
 420
 421         /* State should be equivalent to EXIT */
 422         if (policy->governor_data)
 423                 return -EBUSY;
 424
 425         sg_policy = sugov_policy_alloc(policy);
 426         if (!sg_policy)
 427                 return -ENOMEM;
 428
 429         mutex_lock(&global_tunables_lock);
 430
 431         if (global_tunables) {
 432                 if (WARN_ON(have_governor_per_policy())) {
 433                         ret = -EINVAL;
 434                         goto free_sg_policy;
 435                 }
 436                 policy->governor_data = sg_policy;
 437                 sg_policy->tunables = global_tunables;
 438
 439                 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
 440                 goto out;
 441         }
 442
 443         tunables = sugov_tunables_alloc(sg_policy);
 444         if (!tunables) {
 445                 ret = -ENOMEM;
 446                 goto free_sg_policy;
 447         }
 448
 449         tunables->rate_limit_us = LATENCY_MULTIPLIER;
 450         lat = policy->cpuinfo.transition_latency / NSEC_PER_USEC;
 451         if (lat)
 452                 tunables->rate_limit_us *= lat;
 453
 454         policy->governor_data = sg_policy;
 455         sg_policy->tunables = tunables;
 456
 457         ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
 458                                    get_governor_parent_kobj(policy), "%s",
 459                                    schedutil_gov.name);
 460         if (ret)
 461                 goto fail;
 462
 463  out:
 464         mutex_unlock(&global_tunables_lock);
 465
 466         cpufreq_enable_fast_switch(policy);
 467         return 0;
 468
 469  fail:
 470         policy->governor_data = NULL;
 471         sugov_clear_global_tunables();
 472
 473  free_sg_policy:
 474         mutex_unlock(&global_tunables_lock);
 475
 476         sugov_policy_free(sg_policy);
 477         pr_err("initialization failed (error %d)\n", ret);
 478         return ret;
 479 }
 480
 481 static void sugov_exit(struct cpufreq_policy *policy)
 482 {
 483         struct sugov_policy *sg_policy = policy->governor_data;
 484         struct sugov_tunables *tunables = sg_policy->tunables;
 485         unsigned int count;
 486
 487         cpufreq_disable_fast_switch(policy);
 488
 489         mutex_lock(&global_tunables_lock);
 490
 491         count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
 492         policy->governor_data = NULL;
 493         if (!count)
 494                 sugov_clear_global_tunables();
 495
 496         mutex_unlock(&global_tunables_lock);
 497
 498         sugov_policy_free(sg_policy);
 499 }
 500
 501 static int sugov_start(struct cpufreq_policy *policy)
 502 {
 503         struct sugov_policy *sg_policy = policy->governor_data;
 504         unsigned int cpu;
 505
 506         sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
 507         sg_policy->last_freq_update_time = 0;
 508         sg_policy->next_freq = UINT_MAX;
 509         sg_policy->work_in_progress = false;
 510         sg_policy->need_freq_update = false;
 511         sg_policy->cached_raw_freq = 0;
 512
 513         for_each_cpu(cpu, policy->cpus) {
 514                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 515
 516                 memset(sg_cpu, 0, sizeof(*sg_cpu));
 517                 sg_cpu->sg_policy = sg_policy;
 518                 sg_cpu->flags = SCHED_CPUFREQ_RT;
 519                 sg_cpu->iowait_boost_max = policy->cpuinfo.max_freq;
 520                 cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
 521                                              policy_is_shared(policy) ?
 522                                                         sugov_update_shared :
 523                                                         sugov_update_single);
 524         }
 525         return 0;
 526 }
 527
 528 static void sugov_stop(struct cpufreq_policy *policy)
 529 {
 530         struct sugov_policy *sg_policy = policy->governor_data;
 531         unsigned int cpu;
 532
 533         for_each_cpu(cpu, policy->cpus)
 534                 cpufreq_remove_update_util_hook(cpu);
 535
 536         synchronize_sched();
 537
 538         irq_work_sync(&sg_policy->irq_work);
 539         cancel_work_sync(&sg_policy->work);
 540 }
 541
 542 static void sugov_limits(struct cpufreq_policy *policy)
 543 {
 544         struct sugov_policy *sg_policy = policy->governor_data;
 545
 546         if (!policy->fast_switch_enabled) {
 547                 mutex_lock(&sg_policy->work_lock);
 548                 cpufreq_policy_apply_limits(policy);
 549                 mutex_unlock(&sg_policy->work_lock);
 550         }
 551
 552         sg_policy->need_freq_update = true;
 553 }
 554
 555 static struct cpufreq_governor schedutil_gov = {
 556         .name = "schedutil",
 557         .owner = THIS_MODULE,
 558         .init = sugov_init,
 559         .exit = sugov_exit,
 560         .start = sugov_start,
 561         .stop = sugov_stop,
 562         .limits = sugov_limits,
 563 };
 564
 565 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
 566 struct cpufreq_governor *cpufreq_default_governor(void)
 567 {
 568         return &schedutil_gov;
 569 }
 570 #endif
 571
 572 static int __init sugov_register(void)
 573 {
 574         return cpufreq_register_governor(&schedutil_gov);
 575 }
 576 fs_initcall(sugov_register);