drivers/cpufreq/cpufreq_governor.c

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