drivers/cpufreq/cpufreq_governor.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * drivers/cpufreq/cpufreq_governor.c
   4  *
   5  * CPUFREQ governors common code
   6  *
   7  * Copyright    (C) 2001 Russell King
   8  *              (C) 2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
   9  *              (C) 2003 Jun Nakajima <jun.nakajima@intel.com>
  10  *              (C) 2009 Alexander Clouter <alex@digriz.org.uk>
  11  *              (c) 2012 Viresh Kumar <viresh.kumar@linaro.org>
  12  */
  13
  14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  15
  16 #include <linux/export.h>
  17 #include <linux/kernel_stat.h>
  18 #include <linux/slab.h>
  19
  20 #include "cpufreq_governor.h"
  21
  22 #define CPUFREQ_DBS_MIN_SAMPLING_INTERVAL       (2 * TICK_NSEC / NSEC_PER_USEC)
  23
  24 static DEFINE_PER_CPU(struct cpu_dbs_info, cpu_dbs);
  25
  26 static DEFINE_MUTEX(gov_dbs_data_mutex);
  27
  28 /* Common sysfs tunables */
  29 /*
  30  * store_sampling_rate - update sampling rate effective immediately if needed.
  31  *
  32  * If new rate is smaller than the old, simply updating
  33  * dbs.sampling_rate might not be appropriate. For example, if the
  34  * original sampling_rate was 1 second and the requested new sampling rate is 10
  35  * ms because the user needs immediate reaction from ondemand governor, but not
  36  * sure if higher frequency will be required or not, then, the governor may
  37  * change the sampling rate too late; up to 1 second later. Thus, if we are
  38  * reducing the sampling rate, we need to make the new value effective
  39  * immediately.
  40  *
  41  * This must be called with dbs_data->mutex held, otherwise traversing
  42  * policy_dbs_list isn't safe.
  43  */
  44 ssize_t store_sampling_rate(struct gov_attr_set *attr_set, const char *buf,
  45                             size_t count)
  46 {
  47         struct dbs_data *dbs_data = to_dbs_data(attr_set);
  48         struct policy_dbs_info *policy_dbs;
  49         unsigned int sampling_interval;
  50         int ret;
  51
  52         ret = sscanf(buf, "%u", &sampling_interval);
  53         if (ret != 1 || sampling_interval < CPUFREQ_DBS_MIN_SAMPLING_INTERVAL)
  54                 return -EINVAL;
  55
  56         dbs_data->sampling_rate = sampling_interval;
  57
  58         /*
  59          * We are operating under dbs_data->mutex and so the list and its
  60          * entries can't be freed concurrently.
  61          */
  62         list_for_each_entry(policy_dbs, &attr_set->policy_list, list) {
  63                 mutex_lock(&policy_dbs->update_mutex);
  64                 /*
  65                  * On 32-bit architectures this may race with the
  66                  * sample_delay_ns read in dbs_update_util_handler(), but that
  67                  * really doesn't matter.  If the read returns a value that's
  68                  * too big, the sample will be skipped, but the next invocation
  69                  * of dbs_update_util_handler() (when the update has been
  70                  * completed) will take a sample.
  71                  *
  72                  * If this runs in parallel with dbs_work_handler(), we may end
  73                  * up overwriting the sample_delay_ns value that it has just
  74                  * written, but it will be corrected next time a sample is
  75                  * taken, so it shouldn't be significant.
  76                  */
  77                 gov_update_sample_delay(policy_dbs, 0);
  78                 mutex_unlock(&policy_dbs->update_mutex);
  79         }
  80
  81         return count;
  82 }
  83 EXPORT_SYMBOL_GPL(store_sampling_rate);
  84
  85 /**
  86  * gov_update_cpu_data - Update CPU load data.
  87  * @dbs_data: Top-level governor data pointer.
  88  *
  89  * Update CPU load data for all CPUs in the domain governed by @dbs_data
  90  * (that may be a single policy or a bunch of them if governor tunables are
  91  * system-wide).
  92  *
  93  * Call under the @dbs_data mutex.
  94  */
  95 void gov_update_cpu_data(struct dbs_data *dbs_data)
  96 {
  97         struct policy_dbs_info *policy_dbs;
  98
  99         list_for_each_entry(policy_dbs, &dbs_data->attr_set.policy_list, list) {
 100                 unsigned int j;
 101
 102                 for_each_cpu(j, policy_dbs->policy->cpus) {
 103                         struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
 104
 105                         j_cdbs->prev_cpu_idle = get_cpu_idle_time(j, &j_cdbs->prev_update_time,
 106                                                                   dbs_data->io_is_busy);
 107                         if (dbs_data->ignore_nice_load)
 108                                 j_cdbs->prev_cpu_nice = kcpustat_field(&kcpustat_cpu(j), CPUTIME_NICE, j);
 109                 }
 110         }
 111 }
 112 EXPORT_SYMBOL_GPL(gov_update_cpu_data);
 113
 114 unsigned int dbs_update(struct cpufreq_policy *policy)
 115 {
 116         struct policy_dbs_info *policy_dbs = policy->governor_data;
 117         struct dbs_data *dbs_data = policy_dbs->dbs_data;
 118         unsigned int ignore_nice = dbs_data->ignore_nice_load;
 119         unsigned int max_load = 0, idle_periods = UINT_MAX;
 120         unsigned int sampling_rate, io_busy, j;
 121
 122         /*
 123          * Sometimes governors may use an additional multiplier to increase
 124          * sample delays temporarily.  Apply that multiplier to sampling_rate
 125          * so as to keep the wake-up-from-idle detection logic a bit
 126          * conservative.
 127          */
 128         sampling_rate = dbs_data->sampling_rate * policy_dbs->rate_mult;
 129         /*
 130          * For the purpose of ondemand, waiting for disk IO is an indication
 131          * that you're performance critical, and not that the system is actually
 132          * idle, so do not add the iowait time to the CPU idle time then.
 133          */
 134         io_busy = dbs_data->io_is_busy;
 135
 136         /* Get Absolute Load */
 137         for_each_cpu(j, policy->cpus) {
 138                 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
 139                 u64 update_time, cur_idle_time;
 140                 unsigned int idle_time, time_elapsed;
 141                 unsigned int load;
 142
 143                 cur_idle_time = get_cpu_idle_time(j, &update_time, io_busy);
 144
 145                 time_elapsed = update_time - j_cdbs->prev_update_time;
 146                 j_cdbs->prev_update_time = update_time;
 147
 148                 idle_time = cur_idle_time - j_cdbs->prev_cpu_idle;
 149                 j_cdbs->prev_cpu_idle = cur_idle_time;
 150
 151                 if (ignore_nice) {
 152                         u64 cur_nice = kcpustat_field(&kcpustat_cpu(j), CPUTIME_NICE, j);
 153
 154                         idle_time += div_u64(cur_nice - j_cdbs->prev_cpu_nice, NSEC_PER_USEC);
 155                         j_cdbs->prev_cpu_nice = cur_nice;
 156                 }
 157
 158                 if (unlikely(!time_elapsed)) {
 159                         /*
 160                          * That can only happen when this function is called
 161                          * twice in a row with a very short interval between the
 162                          * calls, so the previous load value can be used then.
 163                          */
 164                         load = j_cdbs->prev_load;
 165                 } else if (unlikely((int)idle_time > 2 * sampling_rate &&
 166                                     j_cdbs->prev_load)) {
 167                         /*
 168                          * If the CPU had gone completely idle and a task has
 169                          * just woken up on this CPU now, it would be unfair to
 170                          * calculate 'load' the usual way for this elapsed
 171                          * time-window, because it would show near-zero load,
 172                          * irrespective of how CPU intensive that task actually
 173                          * was. This is undesirable for latency-sensitive bursty
 174                          * workloads.
 175                          *
 176                          * To avoid this, reuse the 'load' from the previous
 177                          * time-window and give this task a chance to start with
 178                          * a reasonably high CPU frequency. However, that
 179                          * shouldn't be over-done, lest we get stuck at a high
 180                          * load (high frequency) for too long, even when the
 181                          * current system load has actually dropped down, so
 182                          * clear prev_load to guarantee that the load will be
 183                          * computed again next time.
 184                          *
 185                          * Detecting this situation is easy: an unusually large
 186                          * 'idle_time' (as compared to the sampling rate)
 187                          * indicates this scenario.
 188                          */
 189                         load = j_cdbs->prev_load;
 190                         j_cdbs->prev_load = 0;
 191                 } else {
 192                         if (time_elapsed >= idle_time) {
 193                                 load = 100 * (time_elapsed - idle_time) / time_elapsed;
 194                         } else {
 195                                 /*
 196                                  * That can happen if idle_time is returned by
 197                                  * get_cpu_idle_time_jiffy().  In that case
 198                                  * idle_time is roughly equal to the difference
 199                                  * between time_elapsed and "busy time" obtained
 200                                  * from CPU statistics.  Then, the "busy time"
 201                                  * can end up being greater than time_elapsed
 202                                  * (for example, if jiffies_64 and the CPU
 203                                  * statistics are updated by different CPUs),
 204                                  * so idle_time may in fact be negative.  That
 205                                  * means, though, that the CPU was busy all
 206                                  * the time (on the rough average) during the
 207                                  * last sampling interval and 100 can be
 208                                  * returned as the load.
 209                                  */
 210                                 load = (int)idle_time < 0 ? 100 : 0;
 211                         }
 212                         j_cdbs->prev_load = load;
 213                 }
 214
 215                 if (unlikely((int)idle_time > 2 * sampling_rate)) {
 216                         unsigned int periods = idle_time / sampling_rate;
 217
 218                         if (periods < idle_periods)
 219                                 idle_periods = periods;
 220                 }
 221
 222                 if (load > max_load)
 223                         max_load = load;
 224         }
 225
 226         policy_dbs->idle_periods = idle_periods;
 227
 228         return max_load;
 229 }
 230 EXPORT_SYMBOL_GPL(dbs_update);
 231
 232 static void dbs_work_handler(struct work_struct *work)
 233 {
 234         struct policy_dbs_info *policy_dbs;
 235         struct cpufreq_policy *policy;
 236         struct dbs_governor *gov;
 237
 238         policy_dbs = container_of(work, struct policy_dbs_info, work);
 239         policy = policy_dbs->policy;
 240         gov = dbs_governor_of(policy);
 241
 242         /*
 243          * Make sure cpufreq_governor_limits() isn't evaluating load or the
 244          * ondemand governor isn't updating the sampling rate in parallel.
 245          */
 246         mutex_lock(&policy_dbs->update_mutex);
 247         gov_update_sample_delay(policy_dbs, gov->gov_dbs_update(policy));
 248         mutex_unlock(&policy_dbs->update_mutex);
 249
 250         /* Allow the utilization update handler to queue up more work. */
 251         atomic_set(&policy_dbs->work_count, 0);
 252         /*
 253          * If the update below is reordered with respect to the sample delay
 254          * modification, the utilization update handler may end up using a stale
 255          * sample delay value.
 256          */
 257         smp_wmb();
 258         policy_dbs->work_in_progress = false;
 259 }
 260
 261 static void dbs_irq_work(struct irq_work *irq_work)
 262 {
 263         struct policy_dbs_info *policy_dbs;
 264
 265         policy_dbs = container_of(irq_work, struct policy_dbs_info, irq_work);
 266         schedule_work_on(smp_processor_id(), &policy_dbs->work);
 267 }
 268
 269 static void dbs_update_util_handler(struct update_util_data *data, u64 time,
 270                                     unsigned int flags)
 271 {
 272         struct cpu_dbs_info *cdbs = container_of(data, struct cpu_dbs_info, update_util);
 273         struct policy_dbs_info *policy_dbs = cdbs->policy_dbs;
 274         u64 delta_ns, lst;
 275
 276         if (!cpufreq_this_cpu_can_update(policy_dbs->policy))
 277                 return;
 278
 279         /*
 280          * The work may not be allowed to be queued up right now.
 281          * Possible reasons:
 282          * - Work has already been queued up or is in progress.
 283          * - It is too early (too little time from the previous sample).
 284          */
 285         if (policy_dbs->work_in_progress)
 286                 return;
 287
 288         /*
 289          * If the reads below are reordered before the check above, the value
 290          * of sample_delay_ns used in the computation may be stale.
 291          */
 292         smp_rmb();
 293         lst = READ_ONCE(policy_dbs->last_sample_time);
 294         delta_ns = time - lst;
 295         if ((s64)delta_ns < policy_dbs->sample_delay_ns)
 296                 return;
 297
 298         /*
 299          * If the policy is not shared, the irq_work may be queued up right away
 300          * at this point.  Otherwise, we need to ensure that only one of the
 301          * CPUs sharing the policy will do that.
 302          */
 303         if (policy_dbs->is_shared) {
 304                 if (!atomic_add_unless(&policy_dbs->work_count, 1, 1))
 305                         return;
 306
 307                 /*
 308                  * If another CPU updated last_sample_time in the meantime, we
 309                  * shouldn't be here, so clear the work counter and bail out.
 310                  */
 311                 if (unlikely(lst != READ_ONCE(policy_dbs->last_sample_time))) {
 312                         atomic_set(&policy_dbs->work_count, 0);
 313                         return;
 314                 }
 315         }
 316
 317         policy_dbs->last_sample_time = time;
 318         policy_dbs->work_in_progress = true;
 319         irq_work_queue(&policy_dbs->irq_work);
 320 }
 321
 322 static void gov_set_update_util(struct policy_dbs_info *policy_dbs,
 323                                 unsigned int delay_us)
 324 {
 325         struct cpufreq_policy *policy = policy_dbs->policy;
 326         int cpu;
 327
 328         gov_update_sample_delay(policy_dbs, delay_us);
 329         policy_dbs->last_sample_time = 0;
 330
 331         for_each_cpu(cpu, policy->cpus) {
 332                 struct cpu_dbs_info *cdbs = &per_cpu(cpu_dbs, cpu);
 333
 334                 cpufreq_add_update_util_hook(cpu, &cdbs->update_util,
 335                                              dbs_update_util_handler);
 336         }
 337 }
 338
 339 static inline void gov_clear_update_util(struct cpufreq_policy *policy)
 340 {
 341         int i;
 342
 343         for_each_cpu(i, policy->cpus)
 344                 cpufreq_remove_update_util_hook(i);
 345
 346         synchronize_rcu();
 347 }
 348
 349 static struct policy_dbs_info *alloc_policy_dbs_info(struct cpufreq_policy *policy,
 350                                                      struct dbs_governor *gov)
 351 {
 352         struct policy_dbs_info *policy_dbs;
 353         int j;
 354
 355         /* Allocate memory for per-policy governor data. */
 356         policy_dbs = gov->alloc();
 357         if (!policy_dbs)
 358                 return NULL;
 359
 360         policy_dbs->policy = policy;
 361         mutex_init(&policy_dbs->update_mutex);
 362         atomic_set(&policy_dbs->work_count, 0);
 363         init_irq_work(&policy_dbs->irq_work, dbs_irq_work);
 364         INIT_WORK(&policy_dbs->work, dbs_work_handler);
 365
 366         /* Set policy_dbs for all CPUs, online+offline */
 367         for_each_cpu(j, policy->related_cpus) {
 368                 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
 369
 370                 j_cdbs->policy_dbs = policy_dbs;
 371         }
 372         return policy_dbs;
 373 }
 374
 375 static void free_policy_dbs_info(struct policy_dbs_info *policy_dbs,
 376                                  struct dbs_governor *gov)
 377 {
 378         int j;
 379
 380         mutex_destroy(&policy_dbs->update_mutex);
 381
 382         for_each_cpu(j, policy_dbs->policy->related_cpus) {
 383                 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
 384
 385                 j_cdbs->policy_dbs = NULL;
 386                 j_cdbs->update_util.func = NULL;
 387         }
 388         gov->free(policy_dbs);
 389 }
 390
 391 static void cpufreq_dbs_data_release(struct kobject *kobj)
 392 {
 393         struct dbs_data *dbs_data = to_dbs_data(to_gov_attr_set(kobj));
 394         struct dbs_governor *gov = dbs_data->gov;
 395
 396         gov->exit(dbs_data);
 397         kfree(dbs_data);
 398 }
 399
 400 int cpufreq_dbs_governor_init(struct cpufreq_policy *policy)
 401 {
 402         struct dbs_governor *gov = dbs_governor_of(policy);
 403         struct dbs_data *dbs_data;
 404         struct policy_dbs_info *policy_dbs;
 405         int ret = 0;
 406
 407         /* State should be equivalent to EXIT */
 408         if (policy->governor_data)
 409                 return -EBUSY;
 410
 411         policy_dbs = alloc_policy_dbs_info(policy, gov);
 412         if (!policy_dbs)
 413                 return -ENOMEM;
 414
 415         /* Protect gov->gdbs_data against concurrent updates. */
 416         mutex_lock(&gov_dbs_data_mutex);
 417
 418         dbs_data = gov->gdbs_data;
 419         if (dbs_data) {
 420                 if (WARN_ON(have_governor_per_policy())) {
 421                         ret = -EINVAL;
 422                         goto free_policy_dbs_info;
 423                 }
 424                 policy_dbs->dbs_data = dbs_data;
 425                 policy->governor_data = policy_dbs;
 426
 427                 gov_attr_set_get(&dbs_data->attr_set, &policy_dbs->list);
 428                 goto out;
 429         }
 430
 431         dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
 432         if (!dbs_data) {
 433                 ret = -ENOMEM;
 434                 goto free_policy_dbs_info;
 435         }
 436
 437         dbs_data->gov = gov;
 438         gov_attr_set_init(&dbs_data->attr_set, &policy_dbs->list);
 439
 440         ret = gov->init(dbs_data);
 441         if (ret)
 442                 goto free_policy_dbs_info;
 443
 444         /*
 445          * The sampling interval should not be less than the transition latency
 446          * of the CPU and it also cannot be too small for dbs_update() to work
 447          * correctly.
 448          */
 449         dbs_data->sampling_rate = max_t(unsigned int,
 450                                         CPUFREQ_DBS_MIN_SAMPLING_INTERVAL,
 451                                         cpufreq_policy_transition_delay_us(policy));
 452
 453         if (!have_governor_per_policy())
 454                 gov->gdbs_data = dbs_data;
 455
 456         policy_dbs->dbs_data = dbs_data;
 457         policy->governor_data = policy_dbs;
 458
 459         gov->kobj_type.sysfs_ops = &governor_sysfs_ops;
 460         gov->kobj_type.release = cpufreq_dbs_data_release;
 461         ret = kobject_init_and_add(&dbs_data->attr_set.kobj, &gov->kobj_type,
 462                                    get_governor_parent_kobj(policy),
 463                                    "%s", gov->gov.name);
 464         if (!ret)
 465                 goto out;
 466
 467         /* Failure, so roll back. */
 468         pr_err("initialization failed (dbs_data kobject init error %d)\n", ret);
 469
 470         kobject_put(&dbs_data->attr_set.kobj);
 471
 472         policy->governor_data = NULL;
 473
 474         if (!have_governor_per_policy())
 475                 gov->gdbs_data = NULL;
 476         gov->exit(dbs_data);
 477         kfree(dbs_data);
 478
 479 free_policy_dbs_info:
 480         free_policy_dbs_info(policy_dbs, gov);
 481
 482 out:
 483         mutex_unlock(&gov_dbs_data_mutex);
 484         return ret;
 485 }
 486 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_init);
 487
 488 void cpufreq_dbs_governor_exit(struct cpufreq_policy *policy)
 489 {
 490         struct dbs_governor *gov = dbs_governor_of(policy);
 491         struct policy_dbs_info *policy_dbs = policy->governor_data;
 492         struct dbs_data *dbs_data = policy_dbs->dbs_data;
 493         unsigned int count;
 494
 495         /* Protect gov->gdbs_data against concurrent updates. */
 496         mutex_lock(&gov_dbs_data_mutex);
 497
 498         count = gov_attr_set_put(&dbs_data->attr_set, &policy_dbs->list);
 499
 500         policy->governor_data = NULL;
 501
 502         if (!count && !have_governor_per_policy())
 503                 gov->gdbs_data = NULL;
 504
 505         free_policy_dbs_info(policy_dbs, gov);
 506
 507         mutex_unlock(&gov_dbs_data_mutex);
 508 }
 509 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_exit);
 510
 511 int cpufreq_dbs_governor_start(struct cpufreq_policy *policy)
 512 {
 513         struct dbs_governor *gov = dbs_governor_of(policy);
 514         struct policy_dbs_info *policy_dbs = policy->governor_data;
 515         struct dbs_data *dbs_data = policy_dbs->dbs_data;
 516         unsigned int sampling_rate, ignore_nice, j;
 517         unsigned int io_busy;
 518
 519         if (!policy->cur)
 520                 return -EINVAL;
 521
 522         policy_dbs->is_shared = policy_is_shared(policy);
 523         policy_dbs->rate_mult = 1;
 524
 525         sampling_rate = dbs_data->sampling_rate;
 526         ignore_nice = dbs_data->ignore_nice_load;
 527         io_busy = dbs_data->io_is_busy;
 528
 529         for_each_cpu(j, policy->cpus) {
 530                 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
 531
 532                 j_cdbs->prev_cpu_idle = get_cpu_idle_time(j, &j_cdbs->prev_update_time, io_busy);
 533                 /*
 534                  * Make the first invocation of dbs_update() compute the load.
 535                  */
 536                 j_cdbs->prev_load = 0;
 537
 538                 if (ignore_nice)
 539                         j_cdbs->prev_cpu_nice = kcpustat_field(&kcpustat_cpu(j), CPUTIME_NICE, j);
 540         }
 541
 542         gov->start(policy);
 543
 544         gov_set_update_util(policy_dbs, sampling_rate);
 545         return 0;
 546 }
 547 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_start);
 548
 549 void cpufreq_dbs_governor_stop(struct cpufreq_policy *policy)
 550 {
 551         struct policy_dbs_info *policy_dbs = policy->governor_data;
 552
 553         gov_clear_update_util(policy_dbs->policy);
 554         irq_work_sync(&policy_dbs->irq_work);
 555         cancel_work_sync(&policy_dbs->work);
 556         atomic_set(&policy_dbs->work_count, 0);
 557         policy_dbs->work_in_progress = false;
 558 }
 559 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_stop);
 560
 561 void cpufreq_dbs_governor_limits(struct cpufreq_policy *policy)
 562 {
 563         struct policy_dbs_info *policy_dbs;
 564
 565         /* Protect gov->gdbs_data against cpufreq_dbs_governor_exit() */
 566         mutex_lock(&gov_dbs_data_mutex);
 567         policy_dbs = policy->governor_data;
 568         if (!policy_dbs)
 569                 goto out;
 570
 571         mutex_lock(&policy_dbs->update_mutex);
 572         cpufreq_policy_apply_limits(policy);
 573         gov_update_sample_delay(policy_dbs, 0);
 574         mutex_unlock(&policy_dbs->update_mutex);
 575
 576 out:
 577         mutex_unlock(&gov_dbs_data_mutex);
 578 }
 579 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_limits);