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/sched.h>
22 #include <linux/slab.h>
24 #include "cpufreq_governor.h"
26 static DEFINE_PER_CPU(struct cpu_dbs_info, cpu_dbs);
28 static DEFINE_MUTEX(gov_dbs_data_mutex);
30 /* Common sysfs tunables */
32 * store_sampling_rate - update sampling rate effective immediately if needed.
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
43 * This must be called with dbs_data->mutex held, otherwise traversing
44 * policy_dbs_list isn't safe.
46 ssize_t store_sampling_rate(struct dbs_data *dbs_data, const char *buf,
49 struct policy_dbs_info *policy_dbs;
52 ret = sscanf(buf, "%u", &rate);
56 dbs_data->sampling_rate = max(rate, dbs_data->min_sampling_rate);
59 * We are operating under dbs_data->mutex and so the list and its
60 * entries can't be freed concurrently.
62 list_for_each_entry(policy_dbs, &dbs_data->policy_dbs_list, list) {
63 mutex_lock(&policy_dbs->timer_mutex);
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.
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.
77 gov_update_sample_delay(policy_dbs, 0);
78 mutex_unlock(&policy_dbs->timer_mutex);
83 EXPORT_SYMBOL_GPL(store_sampling_rate);
86 * gov_update_cpu_data - Update CPU load data.
87 * @dbs_data: Top-level governor data pointer.
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
93 * Call under the @dbs_data mutex.
95 void gov_update_cpu_data(struct dbs_data *dbs_data)
97 struct policy_dbs_info *policy_dbs;
99 list_for_each_entry(policy_dbs, &dbs_data->policy_dbs_list, list) {
102 for_each_cpu(j, policy_dbs->policy->cpus) {
103 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
105 j_cdbs->prev_cpu_idle = get_cpu_idle_time(j, &j_cdbs->prev_cpu_wall,
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];
112 EXPORT_SYMBOL_GPL(gov_update_cpu_data);
114 static inline struct dbs_data *to_dbs_data(struct kobject *kobj)
116 return container_of(kobj, struct dbs_data, kobj);
119 static inline struct governor_attr *to_gov_attr(struct attribute *attr)
121 return container_of(attr, struct governor_attr, attr);
124 static ssize_t governor_show(struct kobject *kobj, struct attribute *attr,
127 struct dbs_data *dbs_data = to_dbs_data(kobj);
128 struct governor_attr *gattr = to_gov_attr(attr);
130 return gattr->show(dbs_data, buf);
133 static ssize_t governor_store(struct kobject *kobj, struct attribute *attr,
134 const char *buf, size_t count)
136 struct dbs_data *dbs_data = to_dbs_data(kobj);
137 struct governor_attr *gattr = to_gov_attr(attr);
140 mutex_lock(&dbs_data->mutex);
142 if (dbs_data->usage_count)
143 ret = gattr->store(dbs_data, buf, count);
145 mutex_unlock(&dbs_data->mutex);
151 * Sysfs Ops for accessing governor attributes.
153 * All show/store invocations for governor specific sysfs attributes, will first
154 * call the below show/store callbacks and the attribute specific callback will
155 * be called from within it.
157 static const struct sysfs_ops governor_sysfs_ops = {
158 .show = governor_show,
159 .store = governor_store,
162 unsigned int dbs_update(struct cpufreq_policy *policy)
164 struct policy_dbs_info *policy_dbs = policy->governor_data;
165 struct dbs_data *dbs_data = policy_dbs->dbs_data;
166 unsigned int ignore_nice = dbs_data->ignore_nice_load;
167 unsigned int max_load = 0;
168 unsigned int sampling_rate, io_busy, j;
171 * Sometimes governors may use an additional multiplier to increase
172 * sample delays temporarily. Apply that multiplier to sampling_rate
173 * so as to keep the wake-up-from-idle detection logic a bit
176 sampling_rate = dbs_data->sampling_rate * policy_dbs->rate_mult;
178 * For the purpose of ondemand, waiting for disk IO is an indication
179 * that you're performance critical, and not that the system is actually
180 * idle, so do not add the iowait time to the CPU idle time then.
182 io_busy = dbs_data->io_is_busy;
184 /* Get Absolute Load */
185 for_each_cpu(j, policy->cpus) {
186 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
187 u64 cur_wall_time, cur_idle_time;
188 unsigned int idle_time, wall_time;
191 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time, io_busy);
193 wall_time = cur_wall_time - j_cdbs->prev_cpu_wall;
194 j_cdbs->prev_cpu_wall = cur_wall_time;
196 idle_time = cur_idle_time - j_cdbs->prev_cpu_idle;
197 j_cdbs->prev_cpu_idle = cur_idle_time;
200 u64 cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
202 idle_time += cputime_to_usecs(cur_nice - j_cdbs->prev_cpu_nice);
203 j_cdbs->prev_cpu_nice = cur_nice;
206 if (unlikely(!wall_time || wall_time < idle_time))
210 * If the CPU had gone completely idle, and a task just woke up
211 * on this CPU now, it would be unfair to calculate 'load' the
212 * usual way for this elapsed time-window, because it will show
213 * near-zero load, irrespective of how CPU intensive that task
214 * actually is. This is undesirable for latency-sensitive bursty
217 * To avoid this, we reuse the 'load' from the previous
218 * time-window and give this task a chance to start with a
219 * reasonably high CPU frequency. (However, we shouldn't over-do
220 * this copy, lest we get stuck at a high load (high frequency)
221 * for too long, even when the current system load has actually
222 * dropped down. So we perform the copy only once, upon the
223 * first wake-up from idle.)
225 * Detecting this situation is easy: the governor's utilization
226 * update handler would not have run during CPU-idle periods.
227 * Hence, an unusually large 'wall_time' (as compared to the
228 * sampling rate) indicates this scenario.
230 * prev_load can be zero in two cases and we must recalculate it
232 * - during long idle intervals
233 * - explicitly set to zero
235 if (unlikely(wall_time > (2 * sampling_rate) &&
236 j_cdbs->prev_load)) {
237 load = j_cdbs->prev_load;
240 * Perform a destructive copy, to ensure that we copy
241 * the previous load only once, upon the first wake-up
244 j_cdbs->prev_load = 0;
246 load = 100 * (wall_time - idle_time) / wall_time;
247 j_cdbs->prev_load = load;
255 EXPORT_SYMBOL_GPL(dbs_update);
257 static void gov_set_update_util(struct policy_dbs_info *policy_dbs,
258 unsigned int delay_us)
260 struct cpufreq_policy *policy = policy_dbs->policy;
263 gov_update_sample_delay(policy_dbs, delay_us);
264 policy_dbs->last_sample_time = 0;
266 for_each_cpu(cpu, policy->cpus) {
267 struct cpu_dbs_info *cdbs = &per_cpu(cpu_dbs, cpu);
269 cpufreq_set_update_util_data(cpu, &cdbs->update_util);
273 static inline void gov_clear_update_util(struct cpufreq_policy *policy)
277 for_each_cpu(i, policy->cpus)
278 cpufreq_set_update_util_data(i, NULL);
283 static void gov_cancel_work(struct cpufreq_policy *policy)
285 struct policy_dbs_info *policy_dbs = policy->governor_data;
287 gov_clear_update_util(policy_dbs->policy);
288 irq_work_sync(&policy_dbs->irq_work);
289 cancel_work_sync(&policy_dbs->work);
290 atomic_set(&policy_dbs->work_count, 0);
291 policy_dbs->work_in_progress = false;
294 static void dbs_work_handler(struct work_struct *work)
296 struct policy_dbs_info *policy_dbs;
297 struct cpufreq_policy *policy;
298 struct dbs_governor *gov;
300 policy_dbs = container_of(work, struct policy_dbs_info, work);
301 policy = policy_dbs->policy;
302 gov = dbs_governor_of(policy);
305 * Make sure cpufreq_governor_limits() isn't evaluating load or the
306 * ondemand governor isn't updating the sampling rate in parallel.
308 mutex_lock(&policy_dbs->timer_mutex);
309 gov_update_sample_delay(policy_dbs, gov->gov_dbs_timer(policy));
310 mutex_unlock(&policy_dbs->timer_mutex);
312 /* Allow the utilization update handler to queue up more work. */
313 atomic_set(&policy_dbs->work_count, 0);
315 * If the update below is reordered with respect to the sample delay
316 * modification, the utilization update handler may end up using a stale
317 * sample delay value.
320 policy_dbs->work_in_progress = false;
323 static void dbs_irq_work(struct irq_work *irq_work)
325 struct policy_dbs_info *policy_dbs;
327 policy_dbs = container_of(irq_work, struct policy_dbs_info, irq_work);
328 schedule_work_on(smp_processor_id(), &policy_dbs->work);
331 static void dbs_update_util_handler(struct update_util_data *data, u64 time,
332 unsigned long util, unsigned long max)
334 struct cpu_dbs_info *cdbs = container_of(data, struct cpu_dbs_info, update_util);
335 struct policy_dbs_info *policy_dbs = cdbs->policy_dbs;
339 * The work may not be allowed to be queued up right now.
341 * - Work has already been queued up or is in progress.
342 * - It is too early (too little time from the previous sample).
344 if (policy_dbs->work_in_progress)
348 * If the reads below are reordered before the check above, the value
349 * of sample_delay_ns used in the computation may be stale.
352 lst = READ_ONCE(policy_dbs->last_sample_time);
353 delta_ns = time - lst;
354 if ((s64)delta_ns < policy_dbs->sample_delay_ns)
358 * If the policy is not shared, the irq_work may be queued up right away
359 * at this point. Otherwise, we need to ensure that only one of the
360 * CPUs sharing the policy will do that.
362 if (policy_dbs->is_shared) {
363 if (!atomic_add_unless(&policy_dbs->work_count, 1, 1))
367 * If another CPU updated last_sample_time in the meantime, we
368 * shouldn't be here, so clear the work counter and bail out.
370 if (unlikely(lst != READ_ONCE(policy_dbs->last_sample_time))) {
371 atomic_set(&policy_dbs->work_count, 0);
376 policy_dbs->last_sample_time = time;
377 policy_dbs->work_in_progress = true;
378 irq_work_queue(&policy_dbs->irq_work);
381 static struct policy_dbs_info *alloc_policy_dbs_info(struct cpufreq_policy *policy,
382 struct dbs_governor *gov)
384 struct policy_dbs_info *policy_dbs;
387 /* Allocate memory for per-policy governor data. */
388 policy_dbs = gov->alloc();
392 policy_dbs->policy = policy;
393 mutex_init(&policy_dbs->timer_mutex);
394 atomic_set(&policy_dbs->work_count, 0);
395 init_irq_work(&policy_dbs->irq_work, dbs_irq_work);
396 INIT_WORK(&policy_dbs->work, dbs_work_handler);
398 /* Set policy_dbs for all CPUs, online+offline */
399 for_each_cpu(j, policy->related_cpus) {
400 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
402 j_cdbs->policy_dbs = policy_dbs;
403 j_cdbs->update_util.func = dbs_update_util_handler;
408 static void free_policy_dbs_info(struct policy_dbs_info *policy_dbs,
409 struct dbs_governor *gov)
413 mutex_destroy(&policy_dbs->timer_mutex);
415 for_each_cpu(j, policy_dbs->policy->related_cpus) {
416 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
418 j_cdbs->policy_dbs = NULL;
419 j_cdbs->update_util.func = NULL;
421 gov->free(policy_dbs);
424 static int cpufreq_governor_init(struct cpufreq_policy *policy)
426 struct dbs_governor *gov = dbs_governor_of(policy);
427 struct dbs_data *dbs_data;
428 struct policy_dbs_info *policy_dbs;
429 unsigned int latency;
432 /* State should be equivalent to EXIT */
433 if (policy->governor_data)
436 policy_dbs = alloc_policy_dbs_info(policy, gov);
440 /* Protect gov->gdbs_data against concurrent updates. */
441 mutex_lock(&gov_dbs_data_mutex);
443 dbs_data = gov->gdbs_data;
445 if (WARN_ON(have_governor_per_policy())) {
447 goto free_policy_dbs_info;
449 policy_dbs->dbs_data = dbs_data;
450 policy->governor_data = policy_dbs;
452 mutex_lock(&dbs_data->mutex);
453 dbs_data->usage_count++;
454 list_add(&policy_dbs->list, &dbs_data->policy_dbs_list);
455 mutex_unlock(&dbs_data->mutex);
459 dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
462 goto free_policy_dbs_info;
465 INIT_LIST_HEAD(&dbs_data->policy_dbs_list);
466 mutex_init(&dbs_data->mutex);
468 ret = gov->init(dbs_data, !policy->governor->initialized);
470 goto free_policy_dbs_info;
472 /* policy latency is in ns. Convert it to us first */
473 latency = policy->cpuinfo.transition_latency / 1000;
477 /* Bring kernel and HW constraints together */
478 dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate,
479 MIN_LATENCY_MULTIPLIER * latency);
480 dbs_data->sampling_rate = max(dbs_data->min_sampling_rate,
481 LATENCY_MULTIPLIER * latency);
483 if (!have_governor_per_policy())
484 gov->gdbs_data = dbs_data;
486 policy->governor_data = policy_dbs;
488 policy_dbs->dbs_data = dbs_data;
489 dbs_data->usage_count = 1;
490 list_add(&policy_dbs->list, &dbs_data->policy_dbs_list);
492 gov->kobj_type.sysfs_ops = &governor_sysfs_ops;
493 ret = kobject_init_and_add(&dbs_data->kobj, &gov->kobj_type,
494 get_governor_parent_kobj(policy),
495 "%s", gov->gov.name);
499 /* Failure, so roll back. */
500 pr_err("cpufreq: Governor initialization failed (dbs_data kobject init error %d)\n", ret);
502 policy->governor_data = NULL;
504 if (!have_governor_per_policy())
505 gov->gdbs_data = NULL;
506 gov->exit(dbs_data, !policy->governor->initialized);
509 free_policy_dbs_info:
510 free_policy_dbs_info(policy_dbs, gov);
513 mutex_unlock(&gov_dbs_data_mutex);
517 static int cpufreq_governor_exit(struct cpufreq_policy *policy)
519 struct dbs_governor *gov = dbs_governor_of(policy);
520 struct policy_dbs_info *policy_dbs = policy->governor_data;
521 struct dbs_data *dbs_data = policy_dbs->dbs_data;
524 /* Protect gov->gdbs_data against concurrent updates. */
525 mutex_lock(&gov_dbs_data_mutex);
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_dbs, gov);
549 mutex_unlock(&gov_dbs_data_mutex);
553 static int cpufreq_governor_start(struct cpufreq_policy *policy)
555 struct dbs_governor *gov = dbs_governor_of(policy);
556 struct policy_dbs_info *policy_dbs = policy->governor_data;
557 struct dbs_data *dbs_data = policy_dbs->dbs_data;
558 unsigned int sampling_rate, ignore_nice, j;
559 unsigned int io_busy;
564 policy_dbs->is_shared = policy_is_shared(policy);
565 policy_dbs->rate_mult = 1;
567 sampling_rate = dbs_data->sampling_rate;
568 ignore_nice = dbs_data->ignore_nice_load;
569 io_busy = dbs_data->io_is_busy;
571 for_each_cpu(j, policy->cpus) {
572 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
573 unsigned int prev_load;
575 j_cdbs->prev_cpu_idle = get_cpu_idle_time(j, &j_cdbs->prev_cpu_wall, io_busy);
577 prev_load = j_cdbs->prev_cpu_wall - j_cdbs->prev_cpu_idle;
578 j_cdbs->prev_load = 100 * prev_load / (unsigned int)j_cdbs->prev_cpu_wall;
581 j_cdbs->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
586 gov_set_update_util(policy_dbs, sampling_rate);
590 static int cpufreq_governor_stop(struct cpufreq_policy *policy)
592 gov_cancel_work(policy);
596 static int cpufreq_governor_limits(struct cpufreq_policy *policy)
598 struct policy_dbs_info *policy_dbs = policy->governor_data;
600 mutex_lock(&policy_dbs->timer_mutex);
602 if (policy->max < policy->cur)
603 __cpufreq_driver_target(policy, policy->max, CPUFREQ_RELATION_H);
604 else if (policy->min > policy->cur)
605 __cpufreq_driver_target(policy, policy->min, CPUFREQ_RELATION_L);
607 gov_update_sample_delay(policy_dbs, 0);
609 mutex_unlock(&policy_dbs->timer_mutex);
614 int cpufreq_governor_dbs(struct cpufreq_policy *policy, unsigned int event)
616 if (event == CPUFREQ_GOV_POLICY_INIT) {
617 return cpufreq_governor_init(policy);
618 } else if (policy->governor_data) {
620 case CPUFREQ_GOV_POLICY_EXIT:
621 return cpufreq_governor_exit(policy);
622 case CPUFREQ_GOV_START:
623 return cpufreq_governor_start(policy);
624 case CPUFREQ_GOV_STOP:
625 return cpufreq_governor_stop(policy);
626 case CPUFREQ_GOV_LIMITS:
627 return cpufreq_governor_limits(policy);
632 EXPORT_SYMBOL_GPL(cpufreq_governor_dbs);