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