2 * drivers/cpufreq/cpufreq_governor.c
4 * CPUFREQ governors common code
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>
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.
17 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
19 #include <linux/export.h>
20 #include <linux/kernel_stat.h>
21 #include <linux/slab.h>
23 #include "cpufreq_governor.h"
25 DEFINE_MUTEX(dbs_data_mutex);
26 EXPORT_SYMBOL_GPL(dbs_data_mutex);
28 /* Common sysfs tunables */
30 * store_sampling_rate - update sampling rate effective immediately if needed.
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
41 * On the other hand, if new rate is larger than the old, then we may evaluate
42 * the load too soon, and it might we worth updating sample_delay_ns then as
45 * This must be called with dbs_data->mutex held, otherwise traversing
46 * policy_dbs_list isn't safe.
48 ssize_t store_sampling_rate(struct dbs_data *dbs_data, const char *buf,
51 struct policy_dbs_info *policy_dbs;
54 ret = sscanf(buf, "%u", &rate);
58 dbs_data->sampling_rate = max(rate, dbs_data->min_sampling_rate);
61 * We are operating under dbs_data->mutex and so the list and its
62 * entries can't be freed concurrently.
64 list_for_each_entry(policy_dbs, &dbs_data->policy_dbs_list, list) {
65 mutex_lock(&policy_dbs->timer_mutex);
67 * On 32-bit architectures this may race with the
68 * sample_delay_ns read in dbs_update_util_handler(), but that
69 * really doesn't matter. If the read returns a value that's
70 * too big, the sample will be skipped, but the next invocation
71 * of dbs_update_util_handler() (when the update has been
72 * completed) will take a sample. If the returned value is too
73 * small, the sample will be taken immediately, but that isn't a
74 * problem, as we want the new rate to take effect immediately
77 * If this runs in parallel with dbs_work_handler(), we may end
78 * up overwriting the sample_delay_ns value that it has just
79 * written, but the difference should not be too big and it will
80 * be corrected next time a sample is taken, so it shouldn't be
83 gov_update_sample_delay(policy_dbs, dbs_data->sampling_rate);
84 mutex_unlock(&policy_dbs->timer_mutex);
89 EXPORT_SYMBOL_GPL(store_sampling_rate);
91 static inline struct dbs_data *to_dbs_data(struct kobject *kobj)
93 return container_of(kobj, struct dbs_data, kobj);
96 static inline struct governor_attr *to_gov_attr(struct attribute *attr)
98 return container_of(attr, struct governor_attr, attr);
101 static ssize_t governor_show(struct kobject *kobj, struct attribute *attr,
104 struct dbs_data *dbs_data = to_dbs_data(kobj);
105 struct governor_attr *gattr = to_gov_attr(attr);
109 ret = gattr->show(dbs_data, buf);
114 static ssize_t governor_store(struct kobject *kobj, struct attribute *attr,
115 const char *buf, size_t count)
117 struct dbs_data *dbs_data = to_dbs_data(kobj);
118 struct governor_attr *gattr = to_gov_attr(attr);
121 mutex_lock(&dbs_data->mutex);
124 ret = gattr->store(dbs_data, buf, count);
126 mutex_unlock(&dbs_data->mutex);
132 * Sysfs Ops for accessing governor attributes.
134 * All show/store invocations for governor specific sysfs attributes, will first
135 * call the below show/store callbacks and the attribute specific callback will
136 * be called from within it.
138 static const struct sysfs_ops governor_sysfs_ops = {
139 .show = governor_show,
140 .store = governor_store,
143 void dbs_check_cpu(struct cpufreq_policy *policy)
145 int cpu = policy->cpu;
146 struct dbs_governor *gov = dbs_governor_of(policy);
147 struct policy_dbs_info *policy_dbs = policy->governor_data;
148 struct dbs_data *dbs_data = policy_dbs->dbs_data;
149 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
150 unsigned int sampling_rate = dbs_data->sampling_rate;
151 unsigned int ignore_nice = dbs_data->ignore_nice_load;
152 unsigned int max_load = 0;
155 if (gov->governor == GOV_ONDEMAND) {
156 struct od_cpu_dbs_info_s *od_dbs_info =
157 gov->get_cpu_dbs_info_s(cpu);
160 * Sometimes, the ondemand governor uses an additional
161 * multiplier to give long delays. So apply this multiplier to
162 * the 'sampling_rate', so as to keep the wake-up-from-idle
163 * detection logic a bit conservative.
165 sampling_rate *= od_dbs_info->rate_mult;
169 /* Get Absolute Load */
170 for_each_cpu(j, policy->cpus) {
171 struct cpu_dbs_info *j_cdbs;
172 u64 cur_wall_time, cur_idle_time;
173 unsigned int idle_time, wall_time;
177 j_cdbs = gov->get_cpu_cdbs(j);
180 * For the purpose of ondemand, waiting for disk IO is
181 * an indication that you're performance critical, and
182 * not that the system is actually idle. So do not add
183 * the iowait time to the cpu idle time.
185 if (gov->governor == GOV_ONDEMAND)
186 io_busy = od_tuners->io_is_busy;
187 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time, io_busy);
189 wall_time = (unsigned int)
190 (cur_wall_time - j_cdbs->prev_cpu_wall);
191 j_cdbs->prev_cpu_wall = cur_wall_time;
193 if (cur_idle_time < j_cdbs->prev_cpu_idle)
194 cur_idle_time = j_cdbs->prev_cpu_idle;
196 idle_time = (unsigned int)
197 (cur_idle_time - j_cdbs->prev_cpu_idle);
198 j_cdbs->prev_cpu_idle = cur_idle_time;
201 struct cpu_dbs_info *cdbs = gov->get_cpu_cdbs(cpu);
203 unsigned long cur_nice_jiffies;
205 cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] -
208 * Assumption: nice time between sampling periods will
209 * be less than 2^32 jiffies for 32 bit sys
211 cur_nice_jiffies = (unsigned long)
212 cputime64_to_jiffies64(cur_nice);
214 cdbs->prev_cpu_nice =
215 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
216 idle_time += jiffies_to_usecs(cur_nice_jiffies);
219 if (unlikely(!wall_time || wall_time < idle_time))
223 * If the CPU had gone completely idle, and a task just woke up
224 * on this CPU now, it would be unfair to calculate 'load' the
225 * usual way for this elapsed time-window, because it will show
226 * near-zero load, irrespective of how CPU intensive that task
227 * actually is. This is undesirable for latency-sensitive bursty
230 * To avoid this, we reuse the 'load' from the previous
231 * time-window and give this task a chance to start with a
232 * reasonably high CPU frequency. (However, we shouldn't over-do
233 * this copy, lest we get stuck at a high load (high frequency)
234 * for too long, even when the current system load has actually
235 * dropped down. So we perform the copy only once, upon the
236 * first wake-up from idle.)
238 * Detecting this situation is easy: the governor's utilization
239 * update handler would not have run during CPU-idle periods.
240 * Hence, an unusually large 'wall_time' (as compared to the
241 * sampling rate) indicates this scenario.
243 * prev_load can be zero in two cases and we must recalculate it
245 * - during long idle intervals
246 * - explicitly set to zero
248 if (unlikely(wall_time > (2 * sampling_rate) &&
249 j_cdbs->prev_load)) {
250 load = j_cdbs->prev_load;
253 * Perform a destructive copy, to ensure that we copy
254 * the previous load only once, upon the first wake-up
257 j_cdbs->prev_load = 0;
259 load = 100 * (wall_time - idle_time) / wall_time;
260 j_cdbs->prev_load = load;
267 gov->gov_check_cpu(cpu, max_load);
269 EXPORT_SYMBOL_GPL(dbs_check_cpu);
271 void gov_set_update_util(struct policy_dbs_info *policy_dbs,
272 unsigned int delay_us)
274 struct cpufreq_policy *policy = policy_dbs->policy;
275 struct dbs_governor *gov = dbs_governor_of(policy);
278 gov_update_sample_delay(policy_dbs, delay_us);
279 policy_dbs->last_sample_time = 0;
281 for_each_cpu(cpu, policy->cpus) {
282 struct cpu_dbs_info *cdbs = gov->get_cpu_cdbs(cpu);
284 cpufreq_set_update_util_data(cpu, &cdbs->update_util);
287 EXPORT_SYMBOL_GPL(gov_set_update_util);
289 static inline void gov_clear_update_util(struct cpufreq_policy *policy)
293 for_each_cpu(i, policy->cpus)
294 cpufreq_set_update_util_data(i, NULL);
299 static void gov_cancel_work(struct cpufreq_policy *policy)
301 struct policy_dbs_info *policy_dbs = policy->governor_data;
303 /* Tell dbs_update_util_handler() to skip queuing up work items. */
304 atomic_inc(&policy_dbs->work_count);
306 * If dbs_update_util_handler() is already running, it may not notice
307 * the incremented work_count, so wait for it to complete to prevent its
308 * work item from being queued up after the cancel_work_sync() below.
310 gov_clear_update_util(policy_dbs->policy);
311 irq_work_sync(&policy_dbs->irq_work);
312 cancel_work_sync(&policy_dbs->work);
313 atomic_set(&policy_dbs->work_count, 0);
316 static void dbs_work_handler(struct work_struct *work)
318 struct policy_dbs_info *policy_dbs;
319 struct cpufreq_policy *policy;
320 struct dbs_governor *gov;
323 policy_dbs = container_of(work, struct policy_dbs_info, work);
324 policy = policy_dbs->policy;
325 gov = dbs_governor_of(policy);
328 * Make sure cpufreq_governor_limits() isn't evaluating load or the
329 * ondemand governor isn't updating the sampling rate in parallel.
331 mutex_lock(&policy_dbs->timer_mutex);
332 delay = gov->gov_dbs_timer(policy);
333 policy_dbs->sample_delay_ns = jiffies_to_nsecs(delay);
334 mutex_unlock(&policy_dbs->timer_mutex);
337 * If the atomic operation below is reordered with respect to the
338 * sample delay modification, the utilization update handler may end
339 * up using a stale sample delay value.
341 smp_mb__before_atomic();
342 atomic_dec(&policy_dbs->work_count);
345 static void dbs_irq_work(struct irq_work *irq_work)
347 struct policy_dbs_info *policy_dbs;
349 policy_dbs = container_of(irq_work, struct policy_dbs_info, irq_work);
350 schedule_work(&policy_dbs->work);
353 static inline void gov_queue_irq_work(struct policy_dbs_info *policy_dbs)
356 irq_work_queue_on(&policy_dbs->irq_work, smp_processor_id());
358 irq_work_queue(&policy_dbs->irq_work);
362 static void dbs_update_util_handler(struct update_util_data *data, u64 time,
363 unsigned long util, unsigned long max)
365 struct cpu_dbs_info *cdbs = container_of(data, struct cpu_dbs_info, update_util);
366 struct policy_dbs_info *policy_dbs = cdbs->policy_dbs;
369 * The work may not be allowed to be queued up right now.
371 * - Work has already been queued up or is in progress.
372 * - The governor is being stopped.
373 * - It is too early (too little time from the previous sample).
375 if (atomic_inc_return(&policy_dbs->work_count) == 1) {
378 delta_ns = time - policy_dbs->last_sample_time;
379 if ((s64)delta_ns >= policy_dbs->sample_delay_ns) {
380 policy_dbs->last_sample_time = time;
381 gov_queue_irq_work(policy_dbs);
385 atomic_dec(&policy_dbs->work_count);
388 static struct policy_dbs_info *alloc_policy_dbs_info(struct cpufreq_policy *policy,
389 struct dbs_governor *gov)
391 struct policy_dbs_info *policy_dbs;
394 /* Allocate memory for the common information for policy->cpus */
395 policy_dbs = kzalloc(sizeof(*policy_dbs), GFP_KERNEL);
399 policy_dbs->policy = policy;
400 mutex_init(&policy_dbs->timer_mutex);
401 atomic_set(&policy_dbs->work_count, 0);
402 init_irq_work(&policy_dbs->irq_work, dbs_irq_work);
403 INIT_WORK(&policy_dbs->work, dbs_work_handler);
405 /* Set policy_dbs for all CPUs, online+offline */
406 for_each_cpu(j, policy->related_cpus) {
407 struct cpu_dbs_info *j_cdbs = gov->get_cpu_cdbs(j);
409 j_cdbs->policy_dbs = policy_dbs;
410 j_cdbs->update_util.func = dbs_update_util_handler;
415 static void free_policy_dbs_info(struct cpufreq_policy *policy,
416 struct dbs_governor *gov)
418 struct cpu_dbs_info *cdbs = gov->get_cpu_cdbs(policy->cpu);
419 struct policy_dbs_info *policy_dbs = cdbs->policy_dbs;
422 mutex_destroy(&policy_dbs->timer_mutex);
424 for_each_cpu(j, policy->related_cpus) {
425 struct cpu_dbs_info *j_cdbs = gov->get_cpu_cdbs(j);
427 j_cdbs->policy_dbs = NULL;
428 j_cdbs->update_util.func = NULL;
433 static int cpufreq_governor_init(struct cpufreq_policy *policy)
435 struct dbs_governor *gov = dbs_governor_of(policy);
436 struct dbs_data *dbs_data = gov->gdbs_data;
437 struct policy_dbs_info *policy_dbs;
438 unsigned int latency;
441 /* State should be equivalent to EXIT */
442 if (policy->governor_data)
445 policy_dbs = alloc_policy_dbs_info(policy, gov);
450 if (WARN_ON(have_governor_per_policy())) {
452 goto free_policy_dbs_info;
454 policy_dbs->dbs_data = dbs_data;
455 policy->governor_data = policy_dbs;
457 mutex_lock(&dbs_data->mutex);
458 dbs_data->usage_count++;
459 list_add(&policy_dbs->list, &dbs_data->policy_dbs_list);
460 mutex_unlock(&dbs_data->mutex);
465 dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
468 goto free_policy_dbs_info;
471 INIT_LIST_HEAD(&dbs_data->policy_dbs_list);
472 mutex_init(&dbs_data->mutex);
474 ret = gov->init(dbs_data, !policy->governor->initialized);
476 goto free_policy_dbs_info;
478 /* policy latency is in ns. Convert it to us first */
479 latency = policy->cpuinfo.transition_latency / 1000;
483 /* Bring kernel and HW constraints together */
484 dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate,
485 MIN_LATENCY_MULTIPLIER * latency);
486 dbs_data->sampling_rate = max(dbs_data->min_sampling_rate,
487 LATENCY_MULTIPLIER * latency);
489 if (!have_governor_per_policy())
490 gov->gdbs_data = dbs_data;
492 policy->governor_data = policy_dbs;
494 policy_dbs->dbs_data = dbs_data;
495 dbs_data->usage_count = 1;
496 list_add(&policy_dbs->list, &dbs_data->policy_dbs_list);
498 gov->kobj_type.sysfs_ops = &governor_sysfs_ops;
499 ret = kobject_init_and_add(&dbs_data->kobj, &gov->kobj_type,
500 get_governor_parent_kobj(policy),
501 "%s", gov->gov.name);
505 /* Failure, so roll back. */
506 pr_err("cpufreq: Governor initialization failed (dbs_data kobject init error %d)\n", ret);
508 policy->governor_data = NULL;
510 if (!have_governor_per_policy())
511 gov->gdbs_data = NULL;
512 gov->exit(dbs_data, !policy->governor->initialized);
515 free_policy_dbs_info:
516 free_policy_dbs_info(policy, gov);
520 static int cpufreq_governor_exit(struct cpufreq_policy *policy)
522 struct dbs_governor *gov = dbs_governor_of(policy);
523 struct policy_dbs_info *policy_dbs = policy->governor_data;
524 struct dbs_data *dbs_data = policy_dbs->dbs_data;
527 mutex_lock(&dbs_data->mutex);
528 list_del(&policy_dbs->list);
529 count = --dbs_data->usage_count;
530 mutex_unlock(&dbs_data->mutex);
533 kobject_put(&dbs_data->kobj);
535 policy->governor_data = NULL;
537 if (!have_governor_per_policy())
538 gov->gdbs_data = NULL;
540 gov->exit(dbs_data, policy->governor->initialized == 1);
541 mutex_destroy(&dbs_data->mutex);
544 policy->governor_data = NULL;
547 free_policy_dbs_info(policy, gov);
551 static int cpufreq_governor_start(struct cpufreq_policy *policy)
553 struct dbs_governor *gov = dbs_governor_of(policy);
554 struct policy_dbs_info *policy_dbs = policy->governor_data;
555 struct dbs_data *dbs_data = policy_dbs->dbs_data;
556 unsigned int sampling_rate, ignore_nice, j, cpu = policy->cpu;
562 sampling_rate = dbs_data->sampling_rate;
563 ignore_nice = dbs_data->ignore_nice_load;
565 if (gov->governor == GOV_ONDEMAND) {
566 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
568 io_busy = od_tuners->io_is_busy;
571 for_each_cpu(j, policy->cpus) {
572 struct cpu_dbs_info *j_cdbs = gov->get_cpu_cdbs(j);
573 unsigned int prev_load;
575 j_cdbs->prev_cpu_idle =
576 get_cpu_idle_time(j, &j_cdbs->prev_cpu_wall, io_busy);
578 prev_load = (unsigned int)(j_cdbs->prev_cpu_wall -
579 j_cdbs->prev_cpu_idle);
580 j_cdbs->prev_load = 100 * prev_load /
581 (unsigned int)j_cdbs->prev_cpu_wall;
584 j_cdbs->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
587 if (gov->governor == GOV_CONSERVATIVE) {
588 struct cs_cpu_dbs_info_s *cs_dbs_info =
589 gov->get_cpu_dbs_info_s(cpu);
591 cs_dbs_info->down_skip = 0;
592 cs_dbs_info->requested_freq = policy->cur;
594 struct od_ops *od_ops = gov->gov_ops;
595 struct od_cpu_dbs_info_s *od_dbs_info = gov->get_cpu_dbs_info_s(cpu);
597 od_dbs_info->rate_mult = 1;
598 od_dbs_info->sample_type = OD_NORMAL_SAMPLE;
599 od_ops->powersave_bias_init_cpu(cpu);
602 gov_set_update_util(policy_dbs, sampling_rate);
606 static int cpufreq_governor_stop(struct cpufreq_policy *policy)
608 gov_cancel_work(policy);
613 static int cpufreq_governor_limits(struct cpufreq_policy *policy)
615 struct policy_dbs_info *policy_dbs = policy->governor_data;
617 mutex_lock(&policy_dbs->timer_mutex);
618 if (policy->max < policy->cur)
619 __cpufreq_driver_target(policy, policy->max, CPUFREQ_RELATION_H);
620 else if (policy->min > policy->cur)
621 __cpufreq_driver_target(policy, policy->min, CPUFREQ_RELATION_L);
622 dbs_check_cpu(policy);
623 mutex_unlock(&policy_dbs->timer_mutex);
628 int cpufreq_governor_dbs(struct cpufreq_policy *policy, unsigned int event)
632 /* Lock governor to block concurrent initialization of governor */
633 mutex_lock(&dbs_data_mutex);
635 if (event == CPUFREQ_GOV_POLICY_INIT) {
636 ret = cpufreq_governor_init(policy);
637 } else if (policy->governor_data) {
639 case CPUFREQ_GOV_POLICY_EXIT:
640 ret = cpufreq_governor_exit(policy);
642 case CPUFREQ_GOV_START:
643 ret = cpufreq_governor_start(policy);
645 case CPUFREQ_GOV_STOP:
646 ret = cpufreq_governor_stop(policy);
648 case CPUFREQ_GOV_LIMITS:
649 ret = cpufreq_governor_limits(policy);
654 mutex_unlock(&dbs_data_mutex);
657 EXPORT_SYMBOL_GPL(cpufreq_governor_dbs);