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
  36         /* The next fields are only needed if fast switch cannot be used. */
  37         struct irq_work irq_work;
  38         struct work_struct work;
  39         struct mutex work_lock;
  40         bool work_in_progress;
  41
  42         bool need_freq_update;
  43 };
  44
  45 struct sugov_cpu {
  46         struct update_util_data update_util;
  47         struct sugov_policy *sg_policy;
  48
  49         unsigned int cached_raw_freq;
  50
  51         /* The fields below are only needed when sharing a policy. */
  52         unsigned long util;
  53         unsigned long max;
  54         u64 last_update;
  55         unsigned int flags;
  56 };
  57
  58 static DEFINE_PER_CPU(struct sugov_cpu, sugov_cpu);
  59
  60 /************************ Governor internals ***********************/
  61
  62 static bool sugov_should_update_freq(struct sugov_policy *sg_policy, u64 time)
  63 {
  64         s64 delta_ns;
  65
  66         if (sg_policy->work_in_progress)
  67                 return false;
  68
  69         if (unlikely(sg_policy->need_freq_update)) {
  70                 sg_policy->need_freq_update = false;
  71                 /*
  72                  * This happens when limits change, so forget the previous
  73                  * next_freq value and force an update.
  74                  */
  75                 sg_policy->next_freq = UINT_MAX;
  76                 return true;
  77         }
  78
  79         delta_ns = time - sg_policy->last_freq_update_time;
  80         return delta_ns >= sg_policy->freq_update_delay_ns;
  81 }
  82
  83 static void sugov_update_commit(struct sugov_policy *sg_policy, u64 time,
  84                                 unsigned int next_freq)
  85 {
  86         struct cpufreq_policy *policy = sg_policy->policy;
  87
  88         sg_policy->last_freq_update_time = time;
  89
  90         if (policy->fast_switch_enabled) {
  91                 if (sg_policy->next_freq == next_freq) {
  92                         trace_cpu_frequency(policy->cur, smp_processor_id());
  93                         return;
  94                 }
  95                 sg_policy->next_freq = next_freq;
  96                 next_freq = cpufreq_driver_fast_switch(policy, next_freq);
  97                 if (next_freq == CPUFREQ_ENTRY_INVALID)
  98                         return;
  99
 100                 policy->cur = next_freq;
 101                 trace_cpu_frequency(next_freq, smp_processor_id());
 102         } else if (sg_policy->next_freq != next_freq) {
 103                 sg_policy->next_freq = next_freq;
 104                 sg_policy->work_in_progress = true;
 105                 irq_work_queue(&sg_policy->irq_work);
 106         }
 107 }
 108
 109 /**
 110  * get_next_freq - Compute a new frequency for a given cpufreq policy.
 111  * @sg_cpu: schedutil cpu object to compute the new frequency for.
 112  * @util: Current CPU utilization.
 113  * @max: CPU capacity.
 114  *
 115  * If the utilization is frequency-invariant, choose the new frequency to be
 116  * proportional to it, that is
 117  *
 118  * next_freq = C * max_freq * util / max
 119  *
 120  * Otherwise, approximate the would-be frequency-invariant utilization by
 121  * util_raw * (curr_freq / max_freq) which leads to
 122  *
 123  * next_freq = C * curr_freq * util_raw / max
 124  *
 125  * Take C = 1.25 for the frequency tipping point at (util / max) = 0.8.
 126  *
 127  * The lowest driver-supported frequency which is equal or greater than the raw
 128  * next_freq (as calculated above) is returned, subject to policy min/max and
 129  * cpufreq driver limitations.
 130  */
 131 static unsigned int get_next_freq(struct sugov_cpu *sg_cpu, unsigned long util,
 132                                   unsigned long max)
 133 {
 134         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 135         struct cpufreq_policy *policy = sg_policy->policy;
 136         unsigned int freq = arch_scale_freq_invariant() ?
 137                                 policy->cpuinfo.max_freq : policy->cur;
 138
 139         freq = (freq + (freq >> 2)) * util / max;
 140
 141         if (freq == sg_cpu->cached_raw_freq && sg_policy->next_freq != UINT_MAX)
 142                 return sg_policy->next_freq;
 143         sg_cpu->cached_raw_freq = freq;
 144         return cpufreq_driver_resolve_freq(policy, freq);
 145 }
 146
 147 static void sugov_get_util(unsigned long *util, unsigned long *max)
 148 {
 149         struct rq *rq = this_rq();
 150         unsigned long cfs_max;
 151
 152         cfs_max = arch_scale_cpu_capacity(NULL, smp_processor_id());
 153
 154         *util = min(rq->cfs.avg.util_avg, cfs_max);
 155         *max = cfs_max;
 156 }
 157
 158 static void sugov_update_single(struct update_util_data *hook, u64 time,
 159                                 unsigned int flags)
 160 {
 161         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 162         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 163         struct cpufreq_policy *policy = sg_policy->policy;
 164         unsigned long util, max;
 165         unsigned int next_f;
 166
 167         if (!sugov_should_update_freq(sg_policy, time))
 168                 return;
 169
 170         if (flags & SCHED_CPUFREQ_RT_DL) {
 171                 next_f = policy->cpuinfo.max_freq;
 172         } else {
 173                 sugov_get_util(&util, &max);
 174                 next_f = get_next_freq(sg_cpu, util, max);
 175         }
 176         sugov_update_commit(sg_policy, time, next_f);
 177 }
 178
 179 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu,
 180                                            unsigned long util, unsigned long max,
 181                                            unsigned int flags)
 182 {
 183         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 184         struct cpufreq_policy *policy = sg_policy->policy;
 185         unsigned int max_f = policy->cpuinfo.max_freq;
 186         u64 last_freq_update_time = sg_policy->last_freq_update_time;
 187         unsigned int j;
 188
 189         if (flags & SCHED_CPUFREQ_RT_DL)
 190                 return max_f;
 191
 192         for_each_cpu(j, policy->cpus) {
 193                 struct sugov_cpu *j_sg_cpu;
 194                 unsigned long j_util, j_max;
 195                 s64 delta_ns;
 196
 197                 if (j == smp_processor_id())
 198                         continue;
 199
 200                 j_sg_cpu = &per_cpu(sugov_cpu, j);
 201                 /*
 202                  * If the CPU utilization was last updated before the previous
 203                  * frequency update and the time elapsed between the last update
 204                  * of the CPU utilization and the last frequency update is long
 205                  * enough, don't take the CPU into account as it probably is
 206                  * idle now.
 207                  */
 208                 delta_ns = last_freq_update_time - j_sg_cpu->last_update;
 209                 if (delta_ns > TICK_NSEC)
 210                         continue;
 211
 212                 if (j_sg_cpu->flags & SCHED_CPUFREQ_RT_DL)
 213                         return max_f;
 214
 215                 j_util = j_sg_cpu->util;
 216                 j_max = j_sg_cpu->max;
 217                 if (j_util * max > j_max * util) {
 218                         util = j_util;
 219                         max = j_max;
 220                 }
 221         }
 222
 223         return get_next_freq(sg_cpu, util, max);
 224 }
 225
 226 static void sugov_update_shared(struct update_util_data *hook, u64 time,
 227                                 unsigned int flags)
 228 {
 229         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 230         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 231         unsigned long util, max;
 232         unsigned int next_f;
 233
 234         sugov_get_util(&util, &max);
 235
 236         raw_spin_lock(&sg_policy->update_lock);
 237
 238         sg_cpu->util = util;
 239         sg_cpu->max = max;
 240         sg_cpu->flags = flags;
 241         sg_cpu->last_update = time;
 242
 243         if (sugov_should_update_freq(sg_policy, time)) {
 244                 next_f = sugov_next_freq_shared(sg_cpu, util, max, flags);
 245                 sugov_update_commit(sg_policy, time, next_f);
 246         }
 247
 248         raw_spin_unlock(&sg_policy->update_lock);
 249 }
 250
 251 static void sugov_work(struct work_struct *work)
 252 {
 253         struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
 254
 255         mutex_lock(&sg_policy->work_lock);
 256         __cpufreq_driver_target(sg_policy->policy, sg_policy->next_freq,
 257                                 CPUFREQ_RELATION_L);
 258         mutex_unlock(&sg_policy->work_lock);
 259
 260         sg_policy->work_in_progress = false;
 261 }
 262
 263 static void sugov_irq_work(struct irq_work *irq_work)
 264 {
 265         struct sugov_policy *sg_policy;
 266
 267         sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
 268         schedule_work_on(smp_processor_id(), &sg_policy->work);
 269 }
 270
 271 /************************** sysfs interface ************************/
 272
 273 static struct sugov_tunables *global_tunables;
 274 static DEFINE_MUTEX(global_tunables_lock);
 275
 276 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
 277 {
 278         return container_of(attr_set, struct sugov_tunables, attr_set);
 279 }
 280
 281 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
 282 {
 283         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 284
 285         return sprintf(buf, "%u\n", tunables->rate_limit_us);
 286 }
 287
 288 static ssize_t rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf,
 289                                    size_t count)
 290 {
 291         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 292         struct sugov_policy *sg_policy;
 293         unsigned int rate_limit_us;
 294
 295         if (kstrtouint(buf, 10, &rate_limit_us))
 296                 return -EINVAL;
 297
 298         tunables->rate_limit_us = rate_limit_us;
 299
 300         list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
 301                 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
 302
 303         return count;
 304 }
 305
 306 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
 307
 308 static struct attribute *sugov_attributes[] = {
 309         &rate_limit_us.attr,
 310         NULL
 311 };
 312
 313 static struct kobj_type sugov_tunables_ktype = {
 314         .default_attrs = sugov_attributes,
 315         .sysfs_ops = &governor_sysfs_ops,
 316 };
 317
 318 /********************** cpufreq governor interface *********************/
 319
 320 static struct cpufreq_governor schedutil_gov;
 321
 322 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
 323 {
 324         struct sugov_policy *sg_policy;
 325
 326         sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
 327         if (!sg_policy)
 328                 return NULL;
 329
 330         sg_policy->policy = policy;
 331         init_irq_work(&sg_policy->irq_work, sugov_irq_work);
 332         INIT_WORK(&sg_policy->work, sugov_work);
 333         mutex_init(&sg_policy->work_lock);
 334         raw_spin_lock_init(&sg_policy->update_lock);
 335         return sg_policy;
 336 }
 337
 338 static void sugov_policy_free(struct sugov_policy *sg_policy)
 339 {
 340         mutex_destroy(&sg_policy->work_lock);
 341         kfree(sg_policy);
 342 }
 343
 344 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
 345 {
 346         struct sugov_tunables *tunables;
 347
 348         tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
 349         if (tunables) {
 350                 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
 351                 if (!have_governor_per_policy())
 352                         global_tunables = tunables;
 353         }
 354         return tunables;
 355 }
 356
 357 static void sugov_tunables_free(struct sugov_tunables *tunables)
 358 {
 359         if (!have_governor_per_policy())
 360                 global_tunables = NULL;
 361
 362         kfree(tunables);
 363 }
 364
 365 static int sugov_init(struct cpufreq_policy *policy)
 366 {
 367         struct sugov_policy *sg_policy;
 368         struct sugov_tunables *tunables;
 369         unsigned int lat;
 370         int ret = 0;
 371
 372         /* State should be equivalent to EXIT */
 373         if (policy->governor_data)
 374                 return -EBUSY;
 375
 376         sg_policy = sugov_policy_alloc(policy);
 377         if (!sg_policy)
 378                 return -ENOMEM;
 379
 380         mutex_lock(&global_tunables_lock);
 381
 382         if (global_tunables) {
 383                 if (WARN_ON(have_governor_per_policy())) {
 384                         ret = -EINVAL;
 385                         goto free_sg_policy;
 386                 }
 387                 policy->governor_data = sg_policy;
 388                 sg_policy->tunables = global_tunables;
 389
 390                 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
 391                 goto out;
 392         }
 393
 394         tunables = sugov_tunables_alloc(sg_policy);
 395         if (!tunables) {
 396                 ret = -ENOMEM;
 397                 goto free_sg_policy;
 398         }
 399
 400         tunables->rate_limit_us = LATENCY_MULTIPLIER;
 401         lat = policy->cpuinfo.transition_latency / NSEC_PER_USEC;
 402         if (lat)
 403                 tunables->rate_limit_us *= lat;
 404
 405         policy->governor_data = sg_policy;
 406         sg_policy->tunables = tunables;
 407
 408         ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
 409                                    get_governor_parent_kobj(policy), "%s",
 410                                    schedutil_gov.name);
 411         if (ret)
 412                 goto fail;
 413
 414  out:
 415         mutex_unlock(&global_tunables_lock);
 416
 417         cpufreq_enable_fast_switch(policy);
 418         return 0;
 419
 420  fail:
 421         policy->governor_data = NULL;
 422         sugov_tunables_free(tunables);
 423
 424  free_sg_policy:
 425         mutex_unlock(&global_tunables_lock);
 426
 427         sugov_policy_free(sg_policy);
 428         pr_err("initialization failed (error %d)\n", ret);
 429         return ret;
 430 }
 431
 432 static void sugov_exit(struct cpufreq_policy *policy)
 433 {
 434         struct sugov_policy *sg_policy = policy->governor_data;
 435         struct sugov_tunables *tunables = sg_policy->tunables;
 436         unsigned int count;
 437
 438         cpufreq_disable_fast_switch(policy);
 439
 440         mutex_lock(&global_tunables_lock);
 441
 442         count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
 443         policy->governor_data = NULL;
 444         if (!count)
 445                 sugov_tunables_free(tunables);
 446
 447         mutex_unlock(&global_tunables_lock);
 448
 449         sugov_policy_free(sg_policy);
 450 }
 451
 452 static int sugov_start(struct cpufreq_policy *policy)
 453 {
 454         struct sugov_policy *sg_policy = policy->governor_data;
 455         unsigned int cpu;
 456
 457         sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
 458         sg_policy->last_freq_update_time = 0;
 459         sg_policy->next_freq = UINT_MAX;
 460         sg_policy->work_in_progress = false;
 461         sg_policy->need_freq_update = false;
 462
 463         for_each_cpu(cpu, policy->cpus) {
 464                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 465
 466                 sg_cpu->sg_policy = sg_policy;
 467                 if (policy_is_shared(policy)) {
 468                         sg_cpu->util = 0;
 469                         sg_cpu->max = 0;
 470                         sg_cpu->flags = SCHED_CPUFREQ_RT;
 471                         sg_cpu->last_update = 0;
 472                         sg_cpu->cached_raw_freq = 0;
 473                         cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
 474                                                      sugov_update_shared);
 475                 } else {
 476                         cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
 477                                                      sugov_update_single);
 478                 }
 479         }
 480         return 0;
 481 }
 482
 483 static void sugov_stop(struct cpufreq_policy *policy)
 484 {
 485         struct sugov_policy *sg_policy = policy->governor_data;
 486         unsigned int cpu;
 487
 488         for_each_cpu(cpu, policy->cpus)
 489                 cpufreq_remove_update_util_hook(cpu);
 490
 491         synchronize_sched();
 492
 493         irq_work_sync(&sg_policy->irq_work);
 494         cancel_work_sync(&sg_policy->work);
 495 }
 496
 497 static void sugov_limits(struct cpufreq_policy *policy)
 498 {
 499         struct sugov_policy *sg_policy = policy->governor_data;
 500
 501         if (!policy->fast_switch_enabled) {
 502                 mutex_lock(&sg_policy->work_lock);
 503                 cpufreq_policy_apply_limits(policy);
 504                 mutex_unlock(&sg_policy->work_lock);
 505         }
 506
 507         sg_policy->need_freq_update = true;
 508 }
 509
 510 static struct cpufreq_governor schedutil_gov = {
 511         .name = "schedutil",
 512         .owner = THIS_MODULE,
 513         .init = sugov_init,
 514         .exit = sugov_exit,
 515         .start = sugov_start,
 516         .stop = sugov_stop,
 517         .limits = sugov_limits,
 518 };
 519
 520 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
 521 struct cpufreq_governor *cpufreq_default_governor(void)
 522 {
 523         return &schedutil_gov;
 524 }
 525 #endif
 526
 527 static int __init sugov_register(void)
 528 {
 529         return cpufreq_register_governor(&schedutil_gov);
 530 }
 531 fs_initcall(sugov_register);