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 static DEFINE_PER_CPU(struct cpu_dbs_info, cpu_dbs);
  26
  27 static DEFINE_MUTEX(gov_dbs_data_mutex);
  28
  29 /* Common sysfs tunables */
  30 /**
  31  * store_sampling_rate - update sampling rate effective immediately if needed.
  32  *
  33  * If new rate is smaller than the old, simply updating
  34  * dbs.sampling_rate might not be appropriate. For example, if the
  35  * original sampling_rate was 1 second and the requested new sampling rate is 10
  36  * ms because the user needs immediate reaction from ondemand governor, but not
  37  * sure if higher frequency will be required or not, then, the governor may
  38  * change the sampling rate too late; up to 1 second later. Thus, if we are
  39  * reducing the sampling rate, we need to make the new value effective
  40  * immediately.
  41  *
  42  * This must be called with dbs_data->mutex held, otherwise traversing
  43  * policy_dbs_list isn't safe.
  44  */
  45 ssize_t store_sampling_rate(struct gov_attr_set *attr_set, const char *buf,
  46                             size_t count)
  47 {
  48         struct dbs_data *dbs_data = to_dbs_data(attr_set);
  49         struct policy_dbs_info *policy_dbs;
  50         unsigned int rate;
  51         int ret;
  52         ret = sscanf(buf, "%u", &rate);
  53         if (ret != 1)
  54                 return -EINVAL;
  55
  56         dbs_data->sampling_rate = max(rate, dbs_data->min_sampling_rate);
  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_cpu(j).cpustat[CPUTIME_NICE];
 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_cpu(j).cpustat[CPUTIME_NICE];
 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(time_elapsed > 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: the governor's
 186                          * utilization update handler would not have run during
 187                          * CPU-idle periods.  Hence, an unusually large
 188                          * 'time_elapsed' (as compared to the sampling rate)
 189                          * indicates this scenario.
 190                          */
 191                         load = j_cdbs->prev_load;
 192                         j_cdbs->prev_load = 0;
 193                 } else {
 194                         if (time_elapsed >= idle_time) {
 195                                 load = 100 * (time_elapsed - idle_time) / time_elapsed;
 196                         } else {
 197                                 /*
 198                                  * That can happen if idle_time is returned by
 199                                  * get_cpu_idle_time_jiffy().  In that case
 200                                  * idle_time is roughly equal to the difference
 201                                  * between time_elapsed and "busy time" obtained
 202                                  * from CPU statistics.  Then, the "busy time"
 203                                  * can end up being greater than time_elapsed
 204                                  * (for example, if jiffies_64 and the CPU
 205                                  * statistics are updated by different CPUs),
 206                                  * so idle_time may in fact be negative.  That
 207                                  * means, though, that the CPU was busy all
 208                                  * the time (on the rough average) during the
 209                                  * last sampling interval and 100 can be
 210                                  * returned as the load.
 211                                  */
 212                                 load = (int)idle_time < 0 ? 100 : 0;
 213                         }
 214                         j_cdbs->prev_load = load;
 215                 }
 216
 217                 if (time_elapsed > 2 * sampling_rate) {
 218                         unsigned int periods = time_elapsed / sampling_rate;
 219
 220                         if (periods < idle_periods)
 221                                 idle_periods = periods;
 222                 }
 223
 224                 if (load > max_load)
 225                         max_load = load;
 226         }
 227
 228         policy_dbs->idle_periods = idle_periods;
 229
 230         return max_load;
 231 }
 232 EXPORT_SYMBOL_GPL(dbs_update);
 233
 234 static void dbs_work_handler(struct work_struct *work)
 235 {
 236         struct policy_dbs_info *policy_dbs;
 237         struct cpufreq_policy *policy;
 238         struct dbs_governor *gov;
 239
 240         policy_dbs = container_of(work, struct policy_dbs_info, work);
 241         policy = policy_dbs->policy;
 242         gov = dbs_governor_of(policy);
 243
 244         /*
 245          * Make sure cpufreq_governor_limits() isn't evaluating load or the
 246          * ondemand governor isn't updating the sampling rate in parallel.
 247          */
 248         mutex_lock(&policy_dbs->update_mutex);
 249         gov_update_sample_delay(policy_dbs, gov->gov_dbs_update(policy));
 250         mutex_unlock(&policy_dbs->update_mutex);
 251
 252         /* Allow the utilization update handler to queue up more work. */
 253         atomic_set(&policy_dbs->work_count, 0);
 254         /*
 255          * If the update below is reordered with respect to the sample delay
 256          * modification, the utilization update handler may end up using a stale
 257          * sample delay value.
 258          */
 259         smp_wmb();
 260         policy_dbs->work_in_progress = false;
 261 }
 262
 263 static void dbs_irq_work(struct irq_work *irq_work)
 264 {
 265         struct policy_dbs_info *policy_dbs;
 266
 267         policy_dbs = container_of(irq_work, struct policy_dbs_info, irq_work);
 268         schedule_work_on(smp_processor_id(), &policy_dbs->work);
 269 }
 270
 271 static void dbs_update_util_handler(struct update_util_data *data, u64 time,
 272                                     unsigned int flags)
 273 {
 274         struct cpu_dbs_info *cdbs = container_of(data, struct cpu_dbs_info, update_util);
 275         struct policy_dbs_info *policy_dbs = cdbs->policy_dbs;
 276         u64 delta_ns, lst;
 277
 278         /*
 279          * The work may not be allowed to be queued up right now.
 280          * Possible reasons:
 281          * - Work has already been queued up or is in progress.
 282          * - It is too early (too little time from the previous sample).
 283          */
 284         if (policy_dbs->work_in_progress)
 285                 return;
 286
 287         /*
 288          * If the reads below are reordered before the check above, the value
 289          * of sample_delay_ns used in the computation may be stale.
 290          */
 291         smp_rmb();
 292         lst = READ_ONCE(policy_dbs->last_sample_time);
 293         delta_ns = time - lst;
 294         if ((s64)delta_ns < policy_dbs->sample_delay_ns)
 295                 return;
 296
 297         /*
 298          * If the policy is not shared, the irq_work may be queued up right away
 299          * at this point.  Otherwise, we need to ensure that only one of the
 300          * CPUs sharing the policy will do that.
 301          */
 302         if (policy_dbs->is_shared) {
 303                 if (!atomic_add_unless(&policy_dbs->work_count, 1, 1))
 304                         return;
 305
 306                 /*
 307                  * If another CPU updated last_sample_time in the meantime, we
 308                  * shouldn't be here, so clear the work counter and bail out.
 309                  */
 310                 if (unlikely(lst != READ_ONCE(policy_dbs->last_sample_time))) {
 311                         atomic_set(&policy_dbs->work_count, 0);
 312                         return;
 313                 }
 314         }
 315
 316         policy_dbs->last_sample_time = time;
 317         policy_dbs->work_in_progress = true;
 318         irq_work_queue(&policy_dbs->irq_work);
 319 }
 320
 321 static void gov_set_update_util(struct policy_dbs_info *policy_dbs,
 322                                 unsigned int delay_us)
 323 {
 324         struct cpufreq_policy *policy = policy_dbs->policy;
 325         int cpu;
 326
 327         gov_update_sample_delay(policy_dbs, delay_us);
 328         policy_dbs->last_sample_time = 0;
 329
 330         for_each_cpu(cpu, policy->cpus) {
 331                 struct cpu_dbs_info *cdbs = &per_cpu(cpu_dbs, cpu);
 332
 333                 cpufreq_add_update_util_hook(cpu, &cdbs->update_util,
 334                                              dbs_update_util_handler);
 335         }
 336 }
 337
 338 static inline void gov_clear_update_util(struct cpufreq_policy *policy)
 339 {
 340         int i;
 341
 342         for_each_cpu(i, policy->cpus)
 343                 cpufreq_remove_update_util_hook(i);
 344
 345         synchronize_sched();
 346 }
 347
 348 static struct policy_dbs_info *alloc_policy_dbs_info(struct cpufreq_policy *policy,
 349                                                      struct dbs_governor *gov)
 350 {
 351         struct policy_dbs_info *policy_dbs;
 352         int j;
 353
 354         /* Allocate memory for per-policy governor data. */
 355         policy_dbs = gov->alloc();
 356         if (!policy_dbs)
 357                 return NULL;
 358
 359         policy_dbs->policy = policy;
 360         mutex_init(&policy_dbs->update_mutex);
 361         atomic_set(&policy_dbs->work_count, 0);
 362         init_irq_work(&policy_dbs->irq_work, dbs_irq_work);
 363         INIT_WORK(&policy_dbs->work, dbs_work_handler);
 364
 365         /* Set policy_dbs for all CPUs, online+offline */
 366         for_each_cpu(j, policy->related_cpus) {
 367                 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
 368
 369                 j_cdbs->policy_dbs = policy_dbs;
 370         }
 371         return policy_dbs;
 372 }
 373
 374 static void free_policy_dbs_info(struct policy_dbs_info *policy_dbs,
 375                                  struct dbs_governor *gov)
 376 {
 377         int j;
 378
 379         mutex_destroy(&policy_dbs->update_mutex);
 380
 381         for_each_cpu(j, policy_dbs->policy->related_cpus) {
 382                 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
 383
 384                 j_cdbs->policy_dbs = NULL;
 385                 j_cdbs->update_util.func = NULL;
 386         }
 387         gov->free(policy_dbs);
 388 }
 389
 390 int cpufreq_dbs_governor_init(struct cpufreq_policy *policy)
 391 {
 392         struct dbs_governor *gov = dbs_governor_of(policy);
 393         struct dbs_data *dbs_data;
 394         struct policy_dbs_info *policy_dbs;
 395         unsigned int latency;
 396         int ret = 0;
 397
 398         /* State should be equivalent to EXIT */
 399         if (policy->governor_data)
 400                 return -EBUSY;
 401
 402         policy_dbs = alloc_policy_dbs_info(policy, gov);
 403         if (!policy_dbs)
 404                 return -ENOMEM;
 405
 406         /* Protect gov->gdbs_data against concurrent updates. */
 407         mutex_lock(&gov_dbs_data_mutex);
 408
 409         dbs_data = gov->gdbs_data;
 410         if (dbs_data) {
 411                 if (WARN_ON(have_governor_per_policy())) {
 412                         ret = -EINVAL;
 413                         goto free_policy_dbs_info;
 414                 }
 415                 policy_dbs->dbs_data = dbs_data;
 416                 policy->governor_data = policy_dbs;
 417
 418                 gov_attr_set_get(&dbs_data->attr_set, &policy_dbs->list);
 419                 goto out;
 420         }
 421
 422         dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
 423         if (!dbs_data) {
 424                 ret = -ENOMEM;
 425                 goto free_policy_dbs_info;
 426         }
 427
 428         gov_attr_set_init(&dbs_data->attr_set, &policy_dbs->list);
 429
 430         ret = gov->init(dbs_data);
 431         if (ret)
 432                 goto free_policy_dbs_info;
 433
 434         /* policy latency is in ns. Convert it to us first */
 435         latency = policy->cpuinfo.transition_latency / 1000;
 436         if (latency == 0)
 437                 latency = 1;
 438
 439         /* Bring kernel and HW constraints together */
 440         dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate,
 441                                           MIN_LATENCY_MULTIPLIER * latency);
 442         dbs_data->sampling_rate = max(dbs_data->min_sampling_rate,
 443                                       LATENCY_MULTIPLIER * latency);
 444
 445         if (!have_governor_per_policy())
 446                 gov->gdbs_data = dbs_data;
 447
 448         policy_dbs->dbs_data = dbs_data;
 449         policy->governor_data = policy_dbs;
 450
 451         gov->kobj_type.sysfs_ops = &governor_sysfs_ops;
 452         ret = kobject_init_and_add(&dbs_data->attr_set.kobj, &gov->kobj_type,
 453                                    get_governor_parent_kobj(policy),
 454                                    "%s", gov->gov.name);
 455         if (!ret)
 456                 goto out;
 457
 458         /* Failure, so roll back. */
 459         pr_err("initialization failed (dbs_data kobject init error %d)\n", ret);
 460
 461         policy->governor_data = NULL;
 462
 463         if (!have_governor_per_policy())
 464                 gov->gdbs_data = NULL;
 465         gov->exit(dbs_data);
 466         kfree(dbs_data);
 467
 468 free_policy_dbs_info:
 469         free_policy_dbs_info(policy_dbs, gov);
 470
 471 out:
 472         mutex_unlock(&gov_dbs_data_mutex);
 473         return ret;
 474 }
 475 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_init);
 476
 477 void cpufreq_dbs_governor_exit(struct cpufreq_policy *policy)
 478 {
 479         struct dbs_governor *gov = dbs_governor_of(policy);
 480         struct policy_dbs_info *policy_dbs = policy->governor_data;
 481         struct dbs_data *dbs_data = policy_dbs->dbs_data;
 482         unsigned int count;
 483
 484         /* Protect gov->gdbs_data against concurrent updates. */
 485         mutex_lock(&gov_dbs_data_mutex);
 486
 487         count = gov_attr_set_put(&dbs_data->attr_set, &policy_dbs->list);
 488
 489         policy->governor_data = NULL;
 490
 491         if (!count) {
 492                 if (!have_governor_per_policy())
 493                         gov->gdbs_data = NULL;
 494
 495                 gov->exit(dbs_data);
 496                 kfree(dbs_data);
 497         }
 498
 499         free_policy_dbs_info(policy_dbs, gov);
 500
 501         mutex_unlock(&gov_dbs_data_mutex);
 502 }
 503 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_exit);
 504
 505 int cpufreq_dbs_governor_start(struct cpufreq_policy *policy)
 506 {
 507         struct dbs_governor *gov = dbs_governor_of(policy);
 508         struct policy_dbs_info *policy_dbs = policy->governor_data;
 509         struct dbs_data *dbs_data = policy_dbs->dbs_data;
 510         unsigned int sampling_rate, ignore_nice, j;
 511         unsigned int io_busy;
 512
 513         if (!policy->cur)
 514                 return -EINVAL;
 515
 516         policy_dbs->is_shared = policy_is_shared(policy);
 517         policy_dbs->rate_mult = 1;
 518
 519         sampling_rate = dbs_data->sampling_rate;
 520         ignore_nice = dbs_data->ignore_nice_load;
 521         io_busy = dbs_data->io_is_busy;
 522
 523         for_each_cpu(j, policy->cpus) {
 524                 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
 525
 526                 j_cdbs->prev_cpu_idle = get_cpu_idle_time(j, &j_cdbs->prev_update_time, io_busy);
 527                 /*
 528                  * Make the first invocation of dbs_update() compute the load.
 529                  */
 530                 j_cdbs->prev_load = 0;
 531
 532                 if (ignore_nice)
 533                         j_cdbs->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 534         }
 535
 536         gov->start(policy);
 537
 538         gov_set_update_util(policy_dbs, sampling_rate);
 539         return 0;
 540 }
 541 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_start);
 542
 543 void cpufreq_dbs_governor_stop(struct cpufreq_policy *policy)
 544 {
 545         struct policy_dbs_info *policy_dbs = policy->governor_data;
 546
 547         gov_clear_update_util(policy_dbs->policy);
 548         irq_work_sync(&policy_dbs->irq_work);
 549         cancel_work_sync(&policy_dbs->work);
 550         atomic_set(&policy_dbs->work_count, 0);
 551         policy_dbs->work_in_progress = false;
 552 }
 553 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_stop);
 554
 555 void cpufreq_dbs_governor_limits(struct cpufreq_policy *policy)
 556 {
 557         struct policy_dbs_info *policy_dbs = policy->governor_data;
 558
 559         mutex_lock(&policy_dbs->update_mutex);
 560         cpufreq_policy_apply_limits(policy);
 561         gov_update_sample_delay(policy_dbs, 0);
 562
 563         mutex_unlock(&policy_dbs->update_mutex);
 564 }
 565 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_limits);