1 // SPDX-License-Identifier: GPL-2.0-only
3 * drivers/cpufreq/cpufreq_governor.c
5 * CPUFREQ governors common code
7 * Copyright (C) 2001 Russell King
8 * (C) 2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
9 * (C) 2003 Jun Nakajima <jun.nakajima@intel.com>
10 * (C) 2009 Alexander Clouter <alex@digriz.org.uk>
11 * (c) 2012 Viresh Kumar <viresh.kumar@linaro.org>
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
16 #include <linux/export.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/slab.h>
20 #include "cpufreq_governor.h"
22 #define CPUFREQ_DBS_MIN_SAMPLING_INTERVAL (2 * TICK_NSEC / NSEC_PER_USEC)
24 static DEFINE_PER_CPU(struct cpu_dbs_info, cpu_dbs);
26 static DEFINE_MUTEX(gov_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 * This must be called with dbs_data->mutex held, otherwise traversing
42 * policy_dbs_list isn't safe.
44 ssize_t store_sampling_rate(struct gov_attr_set *attr_set, const char *buf,
47 struct dbs_data *dbs_data = to_dbs_data(attr_set);
48 struct policy_dbs_info *policy_dbs;
49 unsigned int sampling_interval;
52 ret = sscanf(buf, "%u", &sampling_interval);
53 if (ret != 1 || sampling_interval < CPUFREQ_DBS_MIN_SAMPLING_INTERVAL)
56 dbs_data->sampling_rate = sampling_interval;
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, &attr_set->policy_list, list) {
63 mutex_lock(&policy_dbs->update_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->update_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->attr_set.policy_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_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];
112 EXPORT_SYMBOL_GPL(gov_update_cpu_data);
114 unsigned int dbs_update(struct cpufreq_policy *policy)
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;
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
128 sampling_rate = dbs_data->sampling_rate * policy_dbs->rate_mult;
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.
134 io_busy = dbs_data->io_is_busy;
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;
143 cur_idle_time = get_cpu_idle_time(j, &update_time, io_busy);
145 time_elapsed = update_time - j_cdbs->prev_update_time;
146 j_cdbs->prev_update_time = update_time;
148 idle_time = cur_idle_time - j_cdbs->prev_cpu_idle;
149 j_cdbs->prev_cpu_idle = cur_idle_time;
152 u64 cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
154 idle_time += div_u64(cur_nice - j_cdbs->prev_cpu_nice, NSEC_PER_USEC);
155 j_cdbs->prev_cpu_nice = cur_nice;
158 if (unlikely(!time_elapsed)) {
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.
164 load = j_cdbs->prev_load;
165 } else if (unlikely((int)idle_time > 2 * sampling_rate &&
166 j_cdbs->prev_load)) {
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
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.
185 * Detecting this situation is easy: an unusually large
186 * 'idle_time' (as compared to the sampling rate)
187 * indicates this scenario.
189 load = j_cdbs->prev_load;
190 j_cdbs->prev_load = 0;
192 if (time_elapsed >= idle_time) {
193 load = 100 * (time_elapsed - idle_time) / time_elapsed;
196 * That can happen if idle_time is returned by
197 * get_cpu_idle_time_jiffy(). In that case
198 * idle_time is roughly equal to the difference
199 * between time_elapsed and "busy time" obtained
200 * from CPU statistics. Then, the "busy time"
201 * can end up being greater than time_elapsed
202 * (for example, if jiffies_64 and the CPU
203 * statistics are updated by different CPUs),
204 * so idle_time may in fact be negative. That
205 * means, though, that the CPU was busy all
206 * the time (on the rough average) during the
207 * last sampling interval and 100 can be
208 * returned as the load.
210 load = (int)idle_time < 0 ? 100 : 0;
212 j_cdbs->prev_load = load;
215 if (unlikely((int)idle_time > 2 * sampling_rate)) {
216 unsigned int periods = idle_time / sampling_rate;
218 if (periods < idle_periods)
219 idle_periods = periods;
226 policy_dbs->idle_periods = idle_periods;
230 EXPORT_SYMBOL_GPL(dbs_update);
232 static void dbs_work_handler(struct work_struct *work)
234 struct policy_dbs_info *policy_dbs;
235 struct cpufreq_policy *policy;
236 struct dbs_governor *gov;
238 policy_dbs = container_of(work, struct policy_dbs_info, work);
239 policy = policy_dbs->policy;
240 gov = dbs_governor_of(policy);
243 * Make sure cpufreq_governor_limits() isn't evaluating load or the
244 * ondemand governor isn't updating the sampling rate in parallel.
246 mutex_lock(&policy_dbs->update_mutex);
247 gov_update_sample_delay(policy_dbs, gov->gov_dbs_update(policy));
248 mutex_unlock(&policy_dbs->update_mutex);
250 /* Allow the utilization update handler to queue up more work. */
251 atomic_set(&policy_dbs->work_count, 0);
253 * If the update below is reordered with respect to the sample delay
254 * modification, the utilization update handler may end up using a stale
255 * sample delay value.
258 policy_dbs->work_in_progress = false;
261 static void dbs_irq_work(struct irq_work *irq_work)
263 struct policy_dbs_info *policy_dbs;
265 policy_dbs = container_of(irq_work, struct policy_dbs_info, irq_work);
266 schedule_work_on(smp_processor_id(), &policy_dbs->work);
269 static void dbs_update_util_handler(struct update_util_data *data, u64 time,
272 struct cpu_dbs_info *cdbs = container_of(data, struct cpu_dbs_info, update_util);
273 struct policy_dbs_info *policy_dbs = cdbs->policy_dbs;
276 if (!cpufreq_this_cpu_can_update(policy_dbs->policy))
280 * The work may not be allowed to be queued up right now.
282 * - Work has already been queued up or is in progress.
283 * - It is too early (too little time from the previous sample).
285 if (policy_dbs->work_in_progress)
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.
293 lst = READ_ONCE(policy_dbs->last_sample_time);
294 delta_ns = time - lst;
295 if ((s64)delta_ns < policy_dbs->sample_delay_ns)
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.
303 if (policy_dbs->is_shared) {
304 if (!atomic_add_unless(&policy_dbs->work_count, 1, 1))
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.
311 if (unlikely(lst != READ_ONCE(policy_dbs->last_sample_time))) {
312 atomic_set(&policy_dbs->work_count, 0);
317 policy_dbs->last_sample_time = time;
318 policy_dbs->work_in_progress = true;
319 irq_work_queue(&policy_dbs->irq_work);
322 static void gov_set_update_util(struct policy_dbs_info *policy_dbs,
323 unsigned int delay_us)
325 struct cpufreq_policy *policy = policy_dbs->policy;
328 gov_update_sample_delay(policy_dbs, delay_us);
329 policy_dbs->last_sample_time = 0;
331 for_each_cpu(cpu, policy->cpus) {
332 struct cpu_dbs_info *cdbs = &per_cpu(cpu_dbs, cpu);
334 cpufreq_add_update_util_hook(cpu, &cdbs->update_util,
335 dbs_update_util_handler);
339 static inline void gov_clear_update_util(struct cpufreq_policy *policy)
343 for_each_cpu(i, policy->cpus)
344 cpufreq_remove_update_util_hook(i);
349 static struct policy_dbs_info *alloc_policy_dbs_info(struct cpufreq_policy *policy,
350 struct dbs_governor *gov)
352 struct policy_dbs_info *policy_dbs;
355 /* Allocate memory for per-policy governor data. */
356 policy_dbs = gov->alloc();
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);
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);
370 j_cdbs->policy_dbs = policy_dbs;
375 static void free_policy_dbs_info(struct policy_dbs_info *policy_dbs,
376 struct dbs_governor *gov)
380 mutex_destroy(&policy_dbs->update_mutex);
382 for_each_cpu(j, policy_dbs->policy->related_cpus) {
383 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
385 j_cdbs->policy_dbs = NULL;
386 j_cdbs->update_util.func = NULL;
388 gov->free(policy_dbs);
391 int cpufreq_dbs_governor_init(struct cpufreq_policy *policy)
393 struct dbs_governor *gov = dbs_governor_of(policy);
394 struct dbs_data *dbs_data;
395 struct policy_dbs_info *policy_dbs;
398 /* State should be equivalent to EXIT */
399 if (policy->governor_data)
402 policy_dbs = alloc_policy_dbs_info(policy, gov);
406 /* Protect gov->gdbs_data against concurrent updates. */
407 mutex_lock(&gov_dbs_data_mutex);
409 dbs_data = gov->gdbs_data;
411 if (WARN_ON(have_governor_per_policy())) {
413 goto free_policy_dbs_info;
415 policy_dbs->dbs_data = dbs_data;
416 policy->governor_data = policy_dbs;
418 gov_attr_set_get(&dbs_data->attr_set, &policy_dbs->list);
422 dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
425 goto free_policy_dbs_info;
428 gov_attr_set_init(&dbs_data->attr_set, &policy_dbs->list);
430 ret = gov->init(dbs_data);
432 goto free_policy_dbs_info;
435 * The sampling interval should not be less than the transition latency
436 * of the CPU and it also cannot be too small for dbs_update() to work
439 dbs_data->sampling_rate = max_t(unsigned int,
440 CPUFREQ_DBS_MIN_SAMPLING_INTERVAL,
441 cpufreq_policy_transition_delay_us(policy));
443 if (!have_governor_per_policy())
444 gov->gdbs_data = dbs_data;
446 policy_dbs->dbs_data = dbs_data;
447 policy->governor_data = policy_dbs;
449 gov->kobj_type.sysfs_ops = &governor_sysfs_ops;
450 ret = kobject_init_and_add(&dbs_data->attr_set.kobj, &gov->kobj_type,
451 get_governor_parent_kobj(policy),
452 "%s", gov->gov.name);
456 /* Failure, so roll back. */
457 pr_err("initialization failed (dbs_data kobject init error %d)\n", ret);
459 kobject_put(&dbs_data->attr_set.kobj);
461 policy->governor_data = NULL;
463 if (!have_governor_per_policy())
464 gov->gdbs_data = NULL;
468 free_policy_dbs_info:
469 free_policy_dbs_info(policy_dbs, gov);
472 mutex_unlock(&gov_dbs_data_mutex);
475 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_init);
477 void cpufreq_dbs_governor_exit(struct cpufreq_policy *policy)
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;
484 /* Protect gov->gdbs_data against concurrent updates. */
485 mutex_lock(&gov_dbs_data_mutex);
487 count = gov_attr_set_put(&dbs_data->attr_set, &policy_dbs->list);
489 policy->governor_data = NULL;
492 if (!have_governor_per_policy())
493 gov->gdbs_data = NULL;
499 free_policy_dbs_info(policy_dbs, gov);
501 mutex_unlock(&gov_dbs_data_mutex);
503 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_exit);
505 int cpufreq_dbs_governor_start(struct cpufreq_policy *policy)
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;
516 policy_dbs->is_shared = policy_is_shared(policy);
517 policy_dbs->rate_mult = 1;
519 sampling_rate = dbs_data->sampling_rate;
520 ignore_nice = dbs_data->ignore_nice_load;
521 io_busy = dbs_data->io_is_busy;
523 for_each_cpu(j, policy->cpus) {
524 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
526 j_cdbs->prev_cpu_idle = get_cpu_idle_time(j, &j_cdbs->prev_update_time, io_busy);
528 * Make the first invocation of dbs_update() compute the load.
530 j_cdbs->prev_load = 0;
533 j_cdbs->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
538 gov_set_update_util(policy_dbs, sampling_rate);
541 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_start);
543 void cpufreq_dbs_governor_stop(struct cpufreq_policy *policy)
545 struct policy_dbs_info *policy_dbs = policy->governor_data;
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;
553 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_stop);
555 void cpufreq_dbs_governor_limits(struct cpufreq_policy *policy)
557 struct policy_dbs_info *policy_dbs;
559 /* Protect gov->gdbs_data against cpufreq_dbs_governor_exit() */
560 mutex_lock(&gov_dbs_data_mutex);
561 policy_dbs = policy->governor_data;
565 mutex_lock(&policy_dbs->update_mutex);
566 cpufreq_policy_apply_limits(policy);
567 gov_update_sample_delay(policy_dbs, 0);
568 mutex_unlock(&policy_dbs->update_mutex);
571 mutex_unlock(&gov_dbs_data_mutex);
573 EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_limits);