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 = rq->cpu_capacity_orig;
 151
 152         *util = min(rq->cfs.avg.util_avg, cfs_max);
 153         *max = cfs_max;
 154 }
 155
 156 static void sugov_update_single(struct update_util_data *hook, u64 time,
 157                                 unsigned int flags)
 158 {
 159         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 160         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 161         struct cpufreq_policy *policy = sg_policy->policy;
 162         unsigned long util, max;
 163         unsigned int next_f;
 164
 165         if (!sugov_should_update_freq(sg_policy, time))
 166                 return;
 167
 168         if (flags & SCHED_CPUFREQ_RT_DL) {
 169                 next_f = policy->cpuinfo.max_freq;
 170         } else {
 171                 sugov_get_util(&util, &max);
 172                 next_f = get_next_freq(sg_cpu, util, max);
 173         }
 174         sugov_update_commit(sg_policy, time, next_f);
 175 }
 176
 177 static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu,
 178                                            unsigned long util, unsigned long max,
 179                                            unsigned int flags)
 180 {
 181         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 182         struct cpufreq_policy *policy = sg_policy->policy;
 183         unsigned int max_f = policy->cpuinfo.max_freq;
 184         u64 last_freq_update_time = sg_policy->last_freq_update_time;
 185         unsigned int j;
 186
 187         if (flags & SCHED_CPUFREQ_RT_DL)
 188                 return max_f;
 189
 190         for_each_cpu(j, policy->cpus) {
 191                 struct sugov_cpu *j_sg_cpu;
 192                 unsigned long j_util, j_max;
 193                 s64 delta_ns;
 194
 195                 if (j == smp_processor_id())
 196                         continue;
 197
 198                 j_sg_cpu = &per_cpu(sugov_cpu, j);
 199                 /*
 200                  * If the CPU utilization was last updated before the previous
 201                  * frequency update and the time elapsed between the last update
 202                  * of the CPU utilization and the last frequency update is long
 203                  * enough, don't take the CPU into account as it probably is
 204                  * idle now.
 205                  */
 206                 delta_ns = last_freq_update_time - j_sg_cpu->last_update;
 207                 if (delta_ns > TICK_NSEC)
 208                         continue;
 209
 210                 if (j_sg_cpu->flags & SCHED_CPUFREQ_RT_DL)
 211                         return max_f;
 212
 213                 j_util = j_sg_cpu->util;
 214                 j_max = j_sg_cpu->max;
 215                 if (j_util * max > j_max * util) {
 216                         util = j_util;
 217                         max = j_max;
 218                 }
 219         }
 220
 221         return get_next_freq(sg_cpu, util, max);
 222 }
 223
 224 static void sugov_update_shared(struct update_util_data *hook, u64 time,
 225                                 unsigned int flags)
 226 {
 227         struct sugov_cpu *sg_cpu = container_of(hook, struct sugov_cpu, update_util);
 228         struct sugov_policy *sg_policy = sg_cpu->sg_policy;
 229         unsigned long util, max;
 230         unsigned int next_f;
 231
 232         sugov_get_util(&util, &max);
 233
 234         raw_spin_lock(&sg_policy->update_lock);
 235
 236         sg_cpu->util = util;
 237         sg_cpu->max = max;
 238         sg_cpu->flags = flags;
 239         sg_cpu->last_update = time;
 240
 241         if (sugov_should_update_freq(sg_policy, time)) {
 242                 next_f = sugov_next_freq_shared(sg_cpu, util, max, flags);
 243                 sugov_update_commit(sg_policy, time, next_f);
 244         }
 245
 246         raw_spin_unlock(&sg_policy->update_lock);
 247 }
 248
 249 static void sugov_work(struct work_struct *work)
 250 {
 251         struct sugov_policy *sg_policy = container_of(work, struct sugov_policy, work);
 252
 253         mutex_lock(&sg_policy->work_lock);
 254         __cpufreq_driver_target(sg_policy->policy, sg_policy->next_freq,
 255                                 CPUFREQ_RELATION_L);
 256         mutex_unlock(&sg_policy->work_lock);
 257
 258         sg_policy->work_in_progress = false;
 259 }
 260
 261 static void sugov_irq_work(struct irq_work *irq_work)
 262 {
 263         struct sugov_policy *sg_policy;
 264
 265         sg_policy = container_of(irq_work, struct sugov_policy, irq_work);
 266         schedule_work_on(smp_processor_id(), &sg_policy->work);
 267 }
 268
 269 /************************** sysfs interface ************************/
 270
 271 static struct sugov_tunables *global_tunables;
 272 static DEFINE_MUTEX(global_tunables_lock);
 273
 274 static inline struct sugov_tunables *to_sugov_tunables(struct gov_attr_set *attr_set)
 275 {
 276         return container_of(attr_set, struct sugov_tunables, attr_set);
 277 }
 278
 279 static ssize_t rate_limit_us_show(struct gov_attr_set *attr_set, char *buf)
 280 {
 281         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 282
 283         return sprintf(buf, "%u\n", tunables->rate_limit_us);
 284 }
 285
 286 static ssize_t rate_limit_us_store(struct gov_attr_set *attr_set, const char *buf,
 287                                    size_t count)
 288 {
 289         struct sugov_tunables *tunables = to_sugov_tunables(attr_set);
 290         struct sugov_policy *sg_policy;
 291         unsigned int rate_limit_us;
 292
 293         if (kstrtouint(buf, 10, &rate_limit_us))
 294                 return -EINVAL;
 295
 296         tunables->rate_limit_us = rate_limit_us;
 297
 298         list_for_each_entry(sg_policy, &attr_set->policy_list, tunables_hook)
 299                 sg_policy->freq_update_delay_ns = rate_limit_us * NSEC_PER_USEC;
 300
 301         return count;
 302 }
 303
 304 static struct governor_attr rate_limit_us = __ATTR_RW(rate_limit_us);
 305
 306 static struct attribute *sugov_attributes[] = {
 307         &rate_limit_us.attr,
 308         NULL
 309 };
 310
 311 static struct kobj_type sugov_tunables_ktype = {
 312         .default_attrs = sugov_attributes,
 313         .sysfs_ops = &governor_sysfs_ops,
 314 };
 315
 316 /********************** cpufreq governor interface *********************/
 317
 318 static struct cpufreq_governor schedutil_gov;
 319
 320 static struct sugov_policy *sugov_policy_alloc(struct cpufreq_policy *policy)
 321 {
 322         struct sugov_policy *sg_policy;
 323
 324         sg_policy = kzalloc(sizeof(*sg_policy), GFP_KERNEL);
 325         if (!sg_policy)
 326                 return NULL;
 327
 328         sg_policy->policy = policy;
 329         init_irq_work(&sg_policy->irq_work, sugov_irq_work);
 330         INIT_WORK(&sg_policy->work, sugov_work);
 331         mutex_init(&sg_policy->work_lock);
 332         raw_spin_lock_init(&sg_policy->update_lock);
 333         return sg_policy;
 334 }
 335
 336 static void sugov_policy_free(struct sugov_policy *sg_policy)
 337 {
 338         mutex_destroy(&sg_policy->work_lock);
 339         kfree(sg_policy);
 340 }
 341
 342 static struct sugov_tunables *sugov_tunables_alloc(struct sugov_policy *sg_policy)
 343 {
 344         struct sugov_tunables *tunables;
 345
 346         tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
 347         if (tunables) {
 348                 gov_attr_set_init(&tunables->attr_set, &sg_policy->tunables_hook);
 349                 if (!have_governor_per_policy())
 350                         global_tunables = tunables;
 351         }
 352         return tunables;
 353 }
 354
 355 static void sugov_tunables_free(struct sugov_tunables *tunables)
 356 {
 357         if (!have_governor_per_policy())
 358                 global_tunables = NULL;
 359
 360         kfree(tunables);
 361 }
 362
 363 static int sugov_init(struct cpufreq_policy *policy)
 364 {
 365         struct sugov_policy *sg_policy;
 366         struct sugov_tunables *tunables;
 367         unsigned int lat;
 368         int ret = 0;
 369
 370         /* State should be equivalent to EXIT */
 371         if (policy->governor_data)
 372                 return -EBUSY;
 373
 374         sg_policy = sugov_policy_alloc(policy);
 375         if (!sg_policy)
 376                 return -ENOMEM;
 377
 378         mutex_lock(&global_tunables_lock);
 379
 380         if (global_tunables) {
 381                 if (WARN_ON(have_governor_per_policy())) {
 382                         ret = -EINVAL;
 383                         goto free_sg_policy;
 384                 }
 385                 policy->governor_data = sg_policy;
 386                 sg_policy->tunables = global_tunables;
 387
 388                 gov_attr_set_get(&global_tunables->attr_set, &sg_policy->tunables_hook);
 389                 goto out;
 390         }
 391
 392         tunables = sugov_tunables_alloc(sg_policy);
 393         if (!tunables) {
 394                 ret = -ENOMEM;
 395                 goto free_sg_policy;
 396         }
 397
 398         tunables->rate_limit_us = LATENCY_MULTIPLIER;
 399         lat = policy->cpuinfo.transition_latency / NSEC_PER_USEC;
 400         if (lat)
 401                 tunables->rate_limit_us *= lat;
 402
 403         policy->governor_data = sg_policy;
 404         sg_policy->tunables = tunables;
 405
 406         ret = kobject_init_and_add(&tunables->attr_set.kobj, &sugov_tunables_ktype,
 407                                    get_governor_parent_kobj(policy), "%s",
 408                                    schedutil_gov.name);
 409         if (ret)
 410                 goto fail;
 411
 412  out:
 413         mutex_unlock(&global_tunables_lock);
 414
 415         cpufreq_enable_fast_switch(policy);
 416         return 0;
 417
 418  fail:
 419         policy->governor_data = NULL;
 420         sugov_tunables_free(tunables);
 421
 422  free_sg_policy:
 423         mutex_unlock(&global_tunables_lock);
 424
 425         sugov_policy_free(sg_policy);
 426         pr_err("initialization failed (error %d)\n", ret);
 427         return ret;
 428 }
 429
 430 static void sugov_exit(struct cpufreq_policy *policy)
 431 {
 432         struct sugov_policy *sg_policy = policy->governor_data;
 433         struct sugov_tunables *tunables = sg_policy->tunables;
 434         unsigned int count;
 435
 436         cpufreq_disable_fast_switch(policy);
 437
 438         mutex_lock(&global_tunables_lock);
 439
 440         count = gov_attr_set_put(&tunables->attr_set, &sg_policy->tunables_hook);
 441         policy->governor_data = NULL;
 442         if (!count)
 443                 sugov_tunables_free(tunables);
 444
 445         mutex_unlock(&global_tunables_lock);
 446
 447         sugov_policy_free(sg_policy);
 448 }
 449
 450 static int sugov_start(struct cpufreq_policy *policy)
 451 {
 452         struct sugov_policy *sg_policy = policy->governor_data;
 453         unsigned int cpu;
 454
 455         sg_policy->freq_update_delay_ns = sg_policy->tunables->rate_limit_us * NSEC_PER_USEC;
 456         sg_policy->last_freq_update_time = 0;
 457         sg_policy->next_freq = UINT_MAX;
 458         sg_policy->work_in_progress = false;
 459         sg_policy->need_freq_update = false;
 460
 461         for_each_cpu(cpu, policy->cpus) {
 462                 struct sugov_cpu *sg_cpu = &per_cpu(sugov_cpu, cpu);
 463
 464                 sg_cpu->sg_policy = sg_policy;
 465                 if (policy_is_shared(policy)) {
 466                         sg_cpu->util = 0;
 467                         sg_cpu->max = 0;
 468                         sg_cpu->flags = SCHED_CPUFREQ_RT;
 469                         sg_cpu->last_update = 0;
 470                         sg_cpu->cached_raw_freq = 0;
 471                         cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
 472                                                      sugov_update_shared);
 473                 } else {
 474                         cpufreq_add_update_util_hook(cpu, &sg_cpu->update_util,
 475                                                      sugov_update_single);
 476                 }
 477         }
 478         return 0;
 479 }
 480
 481 static void sugov_stop(struct cpufreq_policy *policy)
 482 {
 483         struct sugov_policy *sg_policy = policy->governor_data;
 484         unsigned int cpu;
 485
 486         for_each_cpu(cpu, policy->cpus)
 487                 cpufreq_remove_update_util_hook(cpu);
 488
 489         synchronize_sched();
 490
 491         irq_work_sync(&sg_policy->irq_work);
 492         cancel_work_sync(&sg_policy->work);
 493 }
 494
 495 static void sugov_limits(struct cpufreq_policy *policy)
 496 {
 497         struct sugov_policy *sg_policy = policy->governor_data;
 498
 499         if (!policy->fast_switch_enabled) {
 500                 mutex_lock(&sg_policy->work_lock);
 501                 cpufreq_policy_apply_limits(policy);
 502                 mutex_unlock(&sg_policy->work_lock);
 503         }
 504
 505         sg_policy->need_freq_update = true;
 506 }
 507
 508 static struct cpufreq_governor schedutil_gov = {
 509         .name = "schedutil",
 510         .owner = THIS_MODULE,
 511         .init = sugov_init,
 512         .exit = sugov_exit,
 513         .start = sugov_start,
 514         .stop = sugov_stop,
 515         .limits = sugov_limits,
 516 };
 517
 518 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_SCHEDUTIL
 519 struct cpufreq_governor *cpufreq_default_governor(void)
 520 {
 521         return &schedutil_gov;
 522 }
 523 #endif
 524
 525 static int __init sugov_register(void)
 526 {
 527         return cpufreq_register_governor(&schedutil_gov);
 528 }
 529 fs_initcall(sugov_register);