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  * sampling_rate_store - 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 sampling_rate_store(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(sampling_rate_store);
  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                 /*
 149                  * cur_idle_time could be smaller than j_cdbs->prev_cpu_idle if
 150                  * it's obtained from get_cpu_idle_time_jiffy() when NOHZ is
 151                  * off, where idle_time is calculated by the difference between
 152                  * time elapsed in jiffies and "busy time" obtained from CPU
 153                  * statistics.  If a CPU is 100% busy, the time elapsed and busy
 154                  * time should grow with the same amount in two consecutive
 155                  * samples, but in practice there could be a tiny difference,
 156                  * making the accumulated idle time decrease sometimes.  Hence,
 157                  * in this case, idle_time should be regarded as 0 in order to
 158                  * make the further process correct.
 159                  */
 160                 if (cur_idle_time > j_cdbs->prev_cpu_idle)
 161                         idle_time = cur_idle_time - j_cdbs->prev_cpu_idle;
 162                 else
 163                         idle_time = 0;
 164
 165                 j_cdbs->prev_cpu_idle = cur_idle_time;
 166
 167                 if (ignore_nice) {
 168                         u64 cur_nice = kcpustat_field(&kcpustat_cpu(j), CPUTIME_NICE, j);
 169
 170                         idle_time += div_u64(cur_nice - j_cdbs->prev_cpu_nice, NSEC_PER_USEC);
 171                         j_cdbs->prev_cpu_nice = cur_nice;
 172                 }
 173
 174                 if (unlikely(!time_elapsed)) {
 175                         /*
 176                          * That can only happen when this function is called
 177                          * twice in a row with a very short interval between the
 178                          * calls, so the previous load value can be used then.
 179                          */
 180                         load = j_cdbs->prev_load;
 181                 } else if (unlikely(idle_time > 2 * sampling_rate &&
 182                                     j_cdbs->prev_load)) {
 183                         /*
 184                          * If the CPU had gone completely idle and a task has
 185                          * just woken up on this CPU now, it would be unfair to
 186                          * calculate 'load' the usual way for this elapsed
 187                          * time-window, because it would show near-zero load,
 188                          * irrespective of how CPU intensive that task actually
 189                          * was. This is undesirable for latency-sensitive bursty
 190                          * workloads.
 191                          *
 192                          * To avoid this, reuse the 'load' from the previous
 193                          * time-window and give this task a chance to start with
 194                          * a reasonably high CPU frequency. However, that
 195                          * shouldn't be over-done, lest we get stuck at a high
 196                          * load (high frequency) for too long, even when the
 197                          * current system load has actually dropped down, so
 198                          * clear prev_load to guarantee that the load will be
 199                          * computed again next time.
 200                          *
 201                          * Detecting this situation is easy: an unusually large
 202                          * 'idle_time' (as compared to the sampling rate)
 203                          * indicates this scenario.
 204                          */
 205                         load = j_cdbs->prev_load;
 206                         j_cdbs->prev_load = 0;
 207                 } else {
 208                         if (time_elapsed > idle_time)
 209                                 load = 100 * (time_elapsed - idle_time) / time_elapsed;
 210                         else
 211                                 load = 0;
 212
 213                         j_cdbs->prev_load = load;
 214                 }
 215
 216                 if (unlikely(idle_time > 2 * sampling_rate)) {
 217                         unsigned int periods = idle_time / sampling_rate;
 218
 219                         if (periods < idle_periods)
 220                                 idle_periods = periods;
 221                 }
 222
 223                 if (load > max_load)
 224                         max_load = load;
 225         }
 226
 227         policy_dbs->idle_periods = idle_periods;
 228
 229         return max_load;
 230 }
 231 EXPORT_SYMBOL_GPL(dbs_update);
 232
 233 static void dbs_work_handler(struct work_struct *work)
 234 {
 235         struct policy_dbs_info *policy_dbs;
 236         struct cpufreq_policy *policy;
 237         struct dbs_governor *gov;
 238
 239         policy_dbs = container_of(work, struct policy_dbs_info, work);
 240         policy = policy_dbs->policy;
 241         gov = dbs_governor_of(policy);
 242
 243         /*
 244          * Make sure cpufreq_governor_limits() isn't evaluating load or the
 245          * ondemand governor isn't updating the sampling rate in parallel.
 246          */
 247         mutex_lock(&policy_dbs->update_mutex);
 248         gov_update_sample_delay(policy_dbs, gov->gov_dbs_update(policy));
 249         mutex_unlock(&policy_dbs->update_mutex);
 250
 251         /* Allow the utilization update handler to queue up more work. */
 252         atomic_set(&policy_dbs->work_count, 0);
 253         /*
 254          * If the update below is reordered with respect to the sample delay
 255          * modification, the utilization update handler may end up using a stale
 256          * sample delay value.
 257          */
 258         smp_wmb();
 259         policy_dbs->work_in_progress = false;
 260 }
 261
 262 static void dbs_irq_work(struct irq_work *irq_work)
 263 {
 264         struct policy_dbs_info *policy_dbs;
 265
 266         policy_dbs = container_of(irq_work, struct policy_dbs_info, irq_work);
 267         schedule_work_on(smp_processor_id(), &policy_dbs->work);
 268 }
 269
 270 static void dbs_update_util_handler(struct update_util_data *data, u64 time,
 271                                     unsigned int flags)
 272 {
 273         struct cpu_dbs_info *cdbs = container_of(data, struct cpu_dbs_info, update_util);
 274         struct policy_dbs_info *policy_dbs = cdbs->policy_dbs;
 275         u64 delta_ns, lst;
 276
 277         if (!cpufreq_this_cpu_can_update(policy_dbs->policy))
 278                 return;
 279
 280         /*
 281          * The work may not be allowed to be queued up right now.
 282          * Possible reasons:
 283          * - Work has already been queued up or is in progress.
 284          * - It is too early (too little time from the previous sample).
 285          */
 286         if (policy_dbs->work_in_progress)
 287                 return;
 288
 289         /*
 290          * If the reads below are reordered before the check above, the value
 291          * of sample_delay_ns used in the computation may be stale.
 292          */
 293         smp_rmb();
 294         lst = READ_ONCE(policy_dbs->last_sample_time);
 295         delta_ns = time - lst;
 296         if ((s64)delta_ns < policy_dbs->sample_delay_ns)
 297                 return;
 298
 299         /*
 300          * If the policy is not shared, the irq_work may be queued up right away
 301          * at this point.  Otherwise, we need to ensure that only one of the
 302          * CPUs sharing the policy will do that.
 303          */
 304         if (policy_dbs->is_shared) {
 305                 if (!atomic_add_unless(&policy_dbs->work_count, 1, 1))
 306                         return;
 307
 308                 /*
 309                  * If another CPU updated last_sample_time in the meantime, we
 310                  * shouldn't be here, so clear the work counter and bail out.
 311                  */
 312                 if (unlikely(lst != READ_ONCE(policy_dbs->last_sample_time))) {
 313                         atomic_set(&policy_dbs->work_count, 0);
 314                         return;
 315                 }
 316         }
 317
 318         policy_dbs->last_sample_time = time;
 319         policy_dbs->work_in_progress = true;
 320         irq_work_queue(&policy_dbs->irq_work);
 321 }
 322
 323 static void gov_set_update_util(struct policy_dbs_info *policy_dbs,
 324                                 unsigned int delay_us)
 325 {
 326         struct cpufreq_policy *policy = policy_dbs->policy;
 327         int cpu;
 328
 329         gov_update_sample_delay(policy_dbs, delay_us);
 330         policy_dbs->last_sample_time = 0;
 331
 332         for_each_cpu(cpu, policy->cpus) {
 333                 struct cpu_dbs_info *cdbs = &per_cpu(cpu_dbs, cpu);
 334
 335                 cpufreq_add_update_util_hook(cpu, &cdbs->update_util,
 336                                              dbs_update_util_handler);
 337         }
 338 }
 339
 340 static inline void gov_clear_update_util(struct cpufreq_policy *policy)
 341 {
 342         int i;
 343
 344         for_each_cpu(i, policy->cpus)
 345                 cpufreq_remove_update_util_hook(i);
 346
 347         synchronize_rcu();
 348 }
 349
 350 static struct policy_dbs_info *alloc_policy_dbs_info(struct cpufreq_policy *policy,
 351                                                      struct dbs_governor *gov)
 352 {
 353         struct policy_dbs_info *policy_dbs;
 354         int j;
 355
 356         /* Allocate memory for per-policy governor data. */
 357         policy_dbs = gov->alloc();
 358         if (!policy_dbs)
 359                 return NULL;
 360
 361         policy_dbs->policy = policy;
 362         mutex_init(&policy_dbs->update_mutex);
 363         atomic_set(&policy_dbs->work_count, 0);
 364         init_irq_work(&policy_dbs->irq_work, dbs_irq_work);
 365         INIT_WORK(&policy_dbs->work, dbs_work_handler);
 366
 367         /* Set policy_dbs for all CPUs, online+offline */
 368         for_each_cpu(j, policy->related_cpus) {
 369                 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
 370
 371                 j_cdbs->policy_dbs = policy_dbs;
 372         }
 373         return policy_dbs;
 374 }
 375
 376 static void free_policy_dbs_info(struct policy_dbs_info *policy_dbs,
 377                                  struct dbs_governor *gov)
 378 {
 379         int j;
 380
 381         mutex_destroy(&policy_dbs->update_mutex);
 382
 383         for_each_cpu(j, policy_dbs->policy->related_cpus) {
 384                 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
 385
 386                 j_cdbs->policy_dbs = NULL;
 387                 j_cdbs->update_util.func = NULL;
 388         }
 389         gov->free(policy_dbs);
 390 }
 391
 392 static void cpufreq_dbs_data_release(struct kobject *kobj)
 393 {
 394         struct dbs_data *dbs_data = to_dbs_data(to_gov_attr_set(kobj));
 395         struct dbs_governor *gov = dbs_data->gov;
 396
 397         gov->exit(dbs_data);
 398         kfree(dbs_data);
 399 }
 400
 401 int cpufreq_dbs_governor_init(struct cpufreq_policy *policy)
 402 {
 403         struct dbs_governor *gov = dbs_governor_of(policy);
 404         struct dbs_data *dbs_data;
 405         struct policy_dbs_info *policy_dbs;
 406         int ret = 0;
 407
 408         /* State should be equivalent to EXIT */
 409         if (policy->governor_data)
 410                 return -EBUSY;
 411
 412         policy_dbs = alloc_policy_dbs_info(policy, gov);
 413         if (!policy_dbs)
 414                 return -ENOMEM;
 415
 416         /* Protect gov->gdbs_data against concurrent updates. */
 417         mutex_lock(&gov_dbs_data_mutex);
 418
 419         dbs_data = gov->gdbs_data;
 420         if (dbs_data) {
 421                 if (WARN_ON(have_governor_per_policy())) {
 422                         ret = -EINVAL;
 423                         goto free_policy_dbs_info;
 424                 }
 425                 policy_dbs->dbs_data = dbs_data;
 426                 policy->governor_data = policy_dbs;
 427
 428                 gov_attr_set_get(&dbs_data->attr_set, &policy_dbs->list);
 429                 goto out;
 430         }
 431
 432         dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
 433         if (!dbs_data) {
 434                 ret = -ENOMEM;
 435                 goto free_policy_dbs_info;
 436         }
 437
 438         dbs_data->gov = gov;
 439         gov_attr_set_init(&dbs_data->attr_set, &policy_dbs->list);
 440
 441         ret = gov->init(dbs_data);
 442         if (ret)
 443                 goto free_dbs_data;
 444
 445         /*
 446          * The sampling interval should not be less than the transition latency
 447          * of the CPU and it also cannot be too small for dbs_update() to work
 448          * correctly.
 449          */
 450         dbs_data->sampling_rate = max_t(unsigned int,
 451                                         CPUFREQ_DBS_MIN_SAMPLING_INTERVAL,
 452                                         cpufreq_policy_transition_delay_us(policy));
 453
 454         if (!have_governor_per_policy())
 455                 gov->gdbs_data = dbs_data;
 456
 457         policy_dbs->dbs_data = dbs_data;
 458         policy->governor_data = policy_dbs;
 459
 460         gov->kobj_type.sysfs_ops = &governor_sysfs_ops;
 461         gov->kobj_type.release = cpufreq_dbs_data_release;
 462         ret = kobject_init_and_add(&dbs_data->attr_set.kobj, &gov->kobj_type,
 463                                    get_governor_parent_kobj(policy),
 464                                    "%s", gov->gov.name);
 465         if (!ret)
 466                 goto out;
 467
 468         /* Failure, so roll back. */
 469         pr_err("initialization failed (dbs_data kobject init error %d)\n", ret);
 470
 471         kobject_put(&dbs_data->attr_set.kobj);
 472
 473         policy->governor_data = NULL;
 474
 475         if (!have_governor_per_policy())
 476                 gov->gdbs_data = NULL;
 477         gov->exit(dbs_data);
 478
 479 free_dbs_data:
 480         kfree(dbs_data);
 481
 482 free_policy_dbs_info:
 483         free_policy_dbs_info(policy_dbs, gov);
 484
 485 out:
 486         mutex_unlock(&gov_dbs_data_mutex);
 487         return ret;
 488 }
 489 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_init);
 490
 491 void cpufreq_dbs_governor_exit(struct cpufreq_policy *policy)
 492 {
 493         struct dbs_governor *gov = dbs_governor_of(policy);
 494         struct policy_dbs_info *policy_dbs = policy->governor_data;
 495         struct dbs_data *dbs_data = policy_dbs->dbs_data;
 496         unsigned int count;
 497
 498         /* Protect gov->gdbs_data against concurrent updates. */
 499         mutex_lock(&gov_dbs_data_mutex);
 500
 501         count = gov_attr_set_put(&dbs_data->attr_set, &policy_dbs->list);
 502
 503         policy->governor_data = NULL;
 504
 505         if (!count && !have_governor_per_policy())
 506                 gov->gdbs_data = NULL;
 507
 508         free_policy_dbs_info(policy_dbs, gov);
 509
 510         mutex_unlock(&gov_dbs_data_mutex);
 511 }
 512 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_exit);
 513
 514 int cpufreq_dbs_governor_start(struct cpufreq_policy *policy)
 515 {
 516         struct dbs_governor *gov = dbs_governor_of(policy);
 517         struct policy_dbs_info *policy_dbs = policy->governor_data;
 518         struct dbs_data *dbs_data = policy_dbs->dbs_data;
 519         unsigned int sampling_rate, ignore_nice, j;
 520         unsigned int io_busy;
 521
 522         if (!policy->cur)
 523                 return -EINVAL;
 524
 525         policy_dbs->is_shared = policy_is_shared(policy);
 526         policy_dbs->rate_mult = 1;
 527
 528         sampling_rate = dbs_data->sampling_rate;
 529         ignore_nice = dbs_data->ignore_nice_load;
 530         io_busy = dbs_data->io_is_busy;
 531
 532         for_each_cpu(j, policy->cpus) {
 533                 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
 534
 535                 j_cdbs->prev_cpu_idle = get_cpu_idle_time(j, &j_cdbs->prev_update_time, io_busy);
 536                 /*
 537                  * Make the first invocation of dbs_update() compute the load.
 538                  */
 539                 j_cdbs->prev_load = 0;
 540
 541                 if (ignore_nice)
 542                         j_cdbs->prev_cpu_nice = kcpustat_field(&kcpustat_cpu(j), CPUTIME_NICE, j);
 543         }
 544
 545         gov->start(policy);
 546
 547         gov_set_update_util(policy_dbs, sampling_rate);
 548         return 0;
 549 }
 550 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_start);
 551
 552 void cpufreq_dbs_governor_stop(struct cpufreq_policy *policy)
 553 {
 554         struct policy_dbs_info *policy_dbs = policy->governor_data;
 555
 556         gov_clear_update_util(policy_dbs->policy);
 557         irq_work_sync(&policy_dbs->irq_work);
 558         cancel_work_sync(&policy_dbs->work);
 559         atomic_set(&policy_dbs->work_count, 0);
 560         policy_dbs->work_in_progress = false;
 561 }
 562 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_stop);
 563
 564 void cpufreq_dbs_governor_limits(struct cpufreq_policy *policy)
 565 {
 566         struct policy_dbs_info *policy_dbs;
 567
 568         /* Protect gov->gdbs_data against cpufreq_dbs_governor_exit() */
 569         mutex_lock(&gov_dbs_data_mutex);
 570         policy_dbs = policy->governor_data;
 571         if (!policy_dbs)
 572                 goto out;
 573
 574         mutex_lock(&policy_dbs->update_mutex);
 575         cpufreq_policy_apply_limits(policy);
 576         gov_update_sample_delay(policy_dbs, 0);
 577         mutex_unlock(&policy_dbs->update_mutex);
 578
 579 out:
 580         mutex_unlock(&gov_dbs_data_mutex);
 581 }
 582 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_limits);