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/sched.h>
  22 #include <linux/slab.h>
  23
  24 #include "cpufreq_governor.h"
  25
  26 static DEFINE_PER_CPU(struct cpu_dbs_info, cpu_dbs);
  27
  28 static DEFINE_MUTEX(gov_dbs_data_mutex);
  29
  30 /* Common sysfs tunables */
  31 /**
  32  * store_sampling_rate - update sampling rate effective immediately if needed.
  33  *
  34  * If new rate is smaller than the old, simply updating
  35  * dbs.sampling_rate might not be appropriate. For example, if the
  36  * original sampling_rate was 1 second and the requested new sampling rate is 10
  37  * ms because the user needs immediate reaction from ondemand governor, but not
  38  * sure if higher frequency will be required or not, then, the governor may
  39  * change the sampling rate too late; up to 1 second later. Thus, if we are
  40  * reducing the sampling rate, we need to make the new value effective
  41  * immediately.
  42  *
  43  * This must be called with dbs_data->mutex held, otherwise traversing
  44  * policy_dbs_list isn't safe.
  45  */
  46 ssize_t store_sampling_rate(struct gov_attr_set *attr_set, const char *buf,
  47                             size_t count)
  48 {
  49         struct dbs_data *dbs_data = to_dbs_data(attr_set);
  50         struct policy_dbs_info *policy_dbs;
  51         unsigned int rate;
  52         int ret;
  53         ret = sscanf(buf, "%u", &rate);
  54         if (ret != 1)
  55                 return -EINVAL;
  56
  57         dbs_data->sampling_rate = max(rate, dbs_data->min_sampling_rate);
  58
  59         /*
  60          * We are operating under dbs_data->mutex and so the list and its
  61          * entries can't be freed concurrently.
  62          */
  63         list_for_each_entry(policy_dbs, &attr_set->policy_list, list) {
  64                 mutex_lock(&policy_dbs->update_mutex);
  65                 /*
  66                  * On 32-bit architectures this may race with the
  67                  * sample_delay_ns read in dbs_update_util_handler(), but that
  68                  * really doesn't matter.  If the read returns a value that's
  69                  * too big, the sample will be skipped, but the next invocation
  70                  * of dbs_update_util_handler() (when the update has been
  71                  * completed) will take a sample.
  72                  *
  73                  * If this runs in parallel with dbs_work_handler(), we may end
  74                  * up overwriting the sample_delay_ns value that it has just
  75                  * written, but it will be corrected next time a sample is
  76                  * taken, so it shouldn't be significant.
  77                  */
  78                 gov_update_sample_delay(policy_dbs, 0);
  79                 mutex_unlock(&policy_dbs->update_mutex);
  80         }
  81
  82         return count;
  83 }
  84 EXPORT_SYMBOL_GPL(store_sampling_rate);
  85
  86 /**
  87  * gov_update_cpu_data - Update CPU load data.
  88  * @dbs_data: Top-level governor data pointer.
  89  *
  90  * Update CPU load data for all CPUs in the domain governed by @dbs_data
  91  * (that may be a single policy or a bunch of them if governor tunables are
  92  * system-wide).
  93  *
  94  * Call under the @dbs_data mutex.
  95  */
  96 void gov_update_cpu_data(struct dbs_data *dbs_data)
  97 {
  98         struct policy_dbs_info *policy_dbs;
  99
 100         list_for_each_entry(policy_dbs, &dbs_data->attr_set.policy_list, list) {
 101                 unsigned int j;
 102
 103                 for_each_cpu(j, policy_dbs->policy->cpus) {
 104                         struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
 105
 106                         j_cdbs->prev_cpu_idle = get_cpu_idle_time(j, &j_cdbs->prev_update_time,
 107                                                                   dbs_data->io_is_busy);
 108                         if (dbs_data->ignore_nice_load)
 109                                 j_cdbs->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 110                 }
 111         }
 112 }
 113 EXPORT_SYMBOL_GPL(gov_update_cpu_data);
 114
 115 unsigned int dbs_update(struct cpufreq_policy *policy)
 116 {
 117         struct policy_dbs_info *policy_dbs = policy->governor_data;
 118         struct dbs_data *dbs_data = policy_dbs->dbs_data;
 119         unsigned int ignore_nice = dbs_data->ignore_nice_load;
 120         unsigned int max_load = 0, idle_periods = UINT_MAX;
 121         unsigned int sampling_rate, io_busy, j;
 122
 123         /*
 124          * Sometimes governors may use an additional multiplier to increase
 125          * sample delays temporarily.  Apply that multiplier to sampling_rate
 126          * so as to keep the wake-up-from-idle detection logic a bit
 127          * conservative.
 128          */
 129         sampling_rate = dbs_data->sampling_rate * policy_dbs->rate_mult;
 130         /*
 131          * For the purpose of ondemand, waiting for disk IO is an indication
 132          * that you're performance critical, and not that the system is actually
 133          * idle, so do not add the iowait time to the CPU idle time then.
 134          */
 135         io_busy = dbs_data->io_is_busy;
 136
 137         /* Get Absolute Load */
 138         for_each_cpu(j, policy->cpus) {
 139                 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
 140                 u64 update_time, cur_idle_time;
 141                 unsigned int idle_time, time_elapsed;
 142                 unsigned int load;
 143
 144                 cur_idle_time = get_cpu_idle_time(j, &update_time, io_busy);
 145
 146                 time_elapsed = update_time - j_cdbs->prev_update_time;
 147                 j_cdbs->prev_update_time = update_time;
 148
 149                 idle_time = cur_idle_time - j_cdbs->prev_cpu_idle;
 150                 j_cdbs->prev_cpu_idle = cur_idle_time;
 151
 152                 if (ignore_nice) {
 153                         u64 cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 154
 155                         idle_time += cputime_to_usecs(cur_nice - j_cdbs->prev_cpu_nice);
 156                         j_cdbs->prev_cpu_nice = cur_nice;
 157                 }
 158
 159                 if (unlikely(!time_elapsed)) {
 160                         /*
 161                          * That can only happen when this function is called
 162                          * twice in a row with a very short interval between the
 163                          * calls, so the previous load value can be used then.
 164                          */
 165                         load = j_cdbs->prev_load;
 166                 } else if (unlikely(time_elapsed > 2 * sampling_rate &&
 167                                     j_cdbs->prev_load)) {
 168                         /*
 169                          * If the CPU had gone completely idle and a task has
 170                          * just woken up on this CPU now, it would be unfair to
 171                          * calculate 'load' the usual way for this elapsed
 172                          * time-window, because it would show near-zero load,
 173                          * irrespective of how CPU intensive that task actually
 174                          * was. This is undesirable for latency-sensitive bursty
 175                          * workloads.
 176                          *
 177                          * To avoid this, reuse the 'load' from the previous
 178                          * time-window and give this task a chance to start with
 179                          * a reasonably high CPU frequency. However, that
 180                          * shouldn't be over-done, lest we get stuck at a high
 181                          * load (high frequency) for too long, even when the
 182                          * current system load has actually dropped down, so
 183                          * clear prev_load to guarantee that the load will be
 184                          * computed again next time.
 185                          *
 186                          * Detecting this situation is easy: the governor's
 187                          * utilization update handler would not have run during
 188                          * CPU-idle periods.  Hence, an unusually large
 189                          * 'time_elapsed' (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 (time_elapsed > 2 * sampling_rate) {
 219                         unsigned int periods = time_elapsed / 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         /*
 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_sched();
 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 int cpufreq_dbs_governor_init(struct cpufreq_policy *policy)
 392 {
 393         struct dbs_governor *gov = dbs_governor_of(policy);
 394         struct dbs_data *dbs_data;
 395         struct policy_dbs_info *policy_dbs;
 396         unsigned int latency;
 397         int ret = 0;
 398
 399         /* State should be equivalent to EXIT */
 400         if (policy->governor_data)
 401                 return -EBUSY;
 402
 403         policy_dbs = alloc_policy_dbs_info(policy, gov);
 404         if (!policy_dbs)
 405                 return -ENOMEM;
 406
 407         /* Protect gov->gdbs_data against concurrent updates. */
 408         mutex_lock(&gov_dbs_data_mutex);
 409
 410         dbs_data = gov->gdbs_data;
 411         if (dbs_data) {
 412                 if (WARN_ON(have_governor_per_policy())) {
 413                         ret = -EINVAL;
 414                         goto free_policy_dbs_info;
 415                 }
 416                 policy_dbs->dbs_data = dbs_data;
 417                 policy->governor_data = policy_dbs;
 418
 419                 gov_attr_set_get(&dbs_data->attr_set, &policy_dbs->list);
 420                 goto out;
 421         }
 422
 423         dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
 424         if (!dbs_data) {
 425                 ret = -ENOMEM;
 426                 goto free_policy_dbs_info;
 427         }
 428
 429         gov_attr_set_init(&dbs_data->attr_set, &policy_dbs->list);
 430
 431         ret = gov->init(dbs_data);
 432         if (ret)
 433                 goto free_policy_dbs_info;
 434
 435         /* policy latency is in ns. Convert it to us first */
 436         latency = policy->cpuinfo.transition_latency / 1000;
 437         if (latency == 0)
 438                 latency = 1;
 439
 440         /* Bring kernel and HW constraints together */
 441         dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate,
 442                                           MIN_LATENCY_MULTIPLIER * latency);
 443         dbs_data->sampling_rate = max(dbs_data->min_sampling_rate,
 444                                       LATENCY_MULTIPLIER * latency);
 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         policy->governor_data = NULL;
 463
 464         if (!have_governor_per_policy())
 465                 gov->gdbs_data = NULL;
 466         gov->exit(dbs_data);
 467         kfree(dbs_data);
 468
 469 free_policy_dbs_info:
 470         free_policy_dbs_info(policy_dbs, gov);
 471
 472 out:
 473         mutex_unlock(&gov_dbs_data_mutex);
 474         return ret;
 475 }
 476 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_init);
 477
 478 void cpufreq_dbs_governor_exit(struct cpufreq_policy *policy)
 479 {
 480         struct dbs_governor *gov = dbs_governor_of(policy);
 481         struct policy_dbs_info *policy_dbs = policy->governor_data;
 482         struct dbs_data *dbs_data = policy_dbs->dbs_data;
 483         unsigned int count;
 484
 485         /* Protect gov->gdbs_data against concurrent updates. */
 486         mutex_lock(&gov_dbs_data_mutex);
 487
 488         count = gov_attr_set_put(&dbs_data->attr_set, &policy_dbs->list);
 489
 490         policy->governor_data = NULL;
 491
 492         if (!count) {
 493                 if (!have_governor_per_policy())
 494                         gov->gdbs_data = NULL;
 495
 496                 gov->exit(dbs_data);
 497                 kfree(dbs_data);
 498         }
 499
 500         free_policy_dbs_info(policy_dbs, gov);
 501
 502         mutex_unlock(&gov_dbs_data_mutex);
 503 }
 504 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_exit);
 505
 506 int cpufreq_dbs_governor_start(struct cpufreq_policy *policy)
 507 {
 508         struct dbs_governor *gov = dbs_governor_of(policy);
 509         struct policy_dbs_info *policy_dbs = policy->governor_data;
 510         struct dbs_data *dbs_data = policy_dbs->dbs_data;
 511         unsigned int sampling_rate, ignore_nice, j;
 512         unsigned int io_busy;
 513
 514         if (!policy->cur)
 515                 return -EINVAL;
 516
 517         policy_dbs->is_shared = policy_is_shared(policy);
 518         policy_dbs->rate_mult = 1;
 519
 520         sampling_rate = dbs_data->sampling_rate;
 521         ignore_nice = dbs_data->ignore_nice_load;
 522         io_busy = dbs_data->io_is_busy;
 523
 524         for_each_cpu(j, policy->cpus) {
 525                 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
 526
 527                 j_cdbs->prev_cpu_idle = get_cpu_idle_time(j, &j_cdbs->prev_update_time, io_busy);
 528                 /*
 529                  * Make the first invocation of dbs_update() compute the load.
 530                  */
 531                 j_cdbs->prev_load = 0;
 532
 533                 if (ignore_nice)
 534                         j_cdbs->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 535         }
 536
 537         gov->start(policy);
 538
 539         gov_set_update_util(policy_dbs, sampling_rate);
 540         return 0;
 541 }
 542 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_start);
 543
 544 void cpufreq_dbs_governor_stop(struct cpufreq_policy *policy)
 545 {
 546         struct policy_dbs_info *policy_dbs = policy->governor_data;
 547
 548         gov_clear_update_util(policy_dbs->policy);
 549         irq_work_sync(&policy_dbs->irq_work);
 550         cancel_work_sync(&policy_dbs->work);
 551         atomic_set(&policy_dbs->work_count, 0);
 552         policy_dbs->work_in_progress = false;
 553 }
 554 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_stop);
 555
 556 void cpufreq_dbs_governor_limits(struct cpufreq_policy *policy)
 557 {
 558         struct policy_dbs_info *policy_dbs = policy->governor_data;
 559
 560         mutex_lock(&policy_dbs->update_mutex);
 561         cpufreq_policy_apply_limits(policy);
 562         gov_update_sample_delay(policy_dbs, 0);
 563
 564         mutex_unlock(&policy_dbs->update_mutex);
 565 }
 566 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_limits);