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 static struct attribute_group *get_sysfs_attr(struct dbs_governor *gov)
30 return have_governor_per_policy() ?
31 gov->attr_group_gov_pol : gov->attr_group_gov_sys;
34 void dbs_check_cpu(struct cpufreq_policy *policy)
36 int cpu = policy->cpu;
37 struct dbs_governor *gov = dbs_governor_of(policy);
38 struct policy_dbs_info *policy_dbs = policy->governor_data;
39 struct dbs_data *dbs_data = policy_dbs->dbs_data;
40 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
41 struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
42 unsigned int sampling_rate;
43 unsigned int max_load = 0;
44 unsigned int ignore_nice;
47 if (gov->governor == GOV_ONDEMAND) {
48 struct od_cpu_dbs_info_s *od_dbs_info =
49 gov->get_cpu_dbs_info_s(cpu);
52 * Sometimes, the ondemand governor uses an additional
53 * multiplier to give long delays. So apply this multiplier to
54 * the 'sampling_rate', so as to keep the wake-up-from-idle
55 * detection logic a bit conservative.
57 sampling_rate = od_tuners->sampling_rate;
58 sampling_rate *= od_dbs_info->rate_mult;
60 ignore_nice = od_tuners->ignore_nice_load;
62 sampling_rate = cs_tuners->sampling_rate;
63 ignore_nice = cs_tuners->ignore_nice_load;
66 /* Get Absolute Load */
67 for_each_cpu(j, policy->cpus) {
68 struct cpu_dbs_info *j_cdbs;
69 u64 cur_wall_time, cur_idle_time;
70 unsigned int idle_time, wall_time;
74 j_cdbs = gov->get_cpu_cdbs(j);
77 * For the purpose of ondemand, waiting for disk IO is
78 * an indication that you're performance critical, and
79 * not that the system is actually idle. So do not add
80 * the iowait time to the cpu idle time.
82 if (gov->governor == GOV_ONDEMAND)
83 io_busy = od_tuners->io_is_busy;
84 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time, io_busy);
86 wall_time = (unsigned int)
87 (cur_wall_time - j_cdbs->prev_cpu_wall);
88 j_cdbs->prev_cpu_wall = cur_wall_time;
90 if (cur_idle_time < j_cdbs->prev_cpu_idle)
91 cur_idle_time = j_cdbs->prev_cpu_idle;
93 idle_time = (unsigned int)
94 (cur_idle_time - j_cdbs->prev_cpu_idle);
95 j_cdbs->prev_cpu_idle = cur_idle_time;
98 struct cpu_dbs_info *cdbs = gov->get_cpu_cdbs(cpu);
100 unsigned long cur_nice_jiffies;
102 cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] -
105 * Assumption: nice time between sampling periods will
106 * be less than 2^32 jiffies for 32 bit sys
108 cur_nice_jiffies = (unsigned long)
109 cputime64_to_jiffies64(cur_nice);
111 cdbs->prev_cpu_nice =
112 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
113 idle_time += jiffies_to_usecs(cur_nice_jiffies);
116 if (unlikely(!wall_time || wall_time < idle_time))
120 * If the CPU had gone completely idle, and a task just woke up
121 * on this CPU now, it would be unfair to calculate 'load' the
122 * usual way for this elapsed time-window, because it will show
123 * near-zero load, irrespective of how CPU intensive that task
124 * actually is. This is undesirable for latency-sensitive bursty
127 * To avoid this, we reuse the 'load' from the previous
128 * time-window and give this task a chance to start with a
129 * reasonably high CPU frequency. (However, we shouldn't over-do
130 * this copy, lest we get stuck at a high load (high frequency)
131 * for too long, even when the current system load has actually
132 * dropped down. So we perform the copy only once, upon the
133 * first wake-up from idle.)
135 * Detecting this situation is easy: the governor's utilization
136 * update handler would not have run during CPU-idle periods.
137 * Hence, an unusually large 'wall_time' (as compared to the
138 * sampling rate) indicates this scenario.
140 * prev_load can be zero in two cases and we must recalculate it
142 * - during long idle intervals
143 * - explicitly set to zero
145 if (unlikely(wall_time > (2 * sampling_rate) &&
146 j_cdbs->prev_load)) {
147 load = j_cdbs->prev_load;
150 * Perform a destructive copy, to ensure that we copy
151 * the previous load only once, upon the first wake-up
154 j_cdbs->prev_load = 0;
156 load = 100 * (wall_time - idle_time) / wall_time;
157 j_cdbs->prev_load = load;
164 gov->gov_check_cpu(cpu, max_load);
166 EXPORT_SYMBOL_GPL(dbs_check_cpu);
168 void gov_set_update_util(struct policy_dbs_info *policy_dbs,
169 unsigned int delay_us)
171 struct cpufreq_policy *policy = policy_dbs->policy;
172 struct dbs_governor *gov = dbs_governor_of(policy);
175 gov_update_sample_delay(policy_dbs, delay_us);
176 policy_dbs->last_sample_time = 0;
178 for_each_cpu(cpu, policy->cpus) {
179 struct cpu_dbs_info *cdbs = gov->get_cpu_cdbs(cpu);
181 cpufreq_set_update_util_data(cpu, &cdbs->update_util);
184 EXPORT_SYMBOL_GPL(gov_set_update_util);
186 static inline void gov_clear_update_util(struct cpufreq_policy *policy)
190 for_each_cpu(i, policy->cpus)
191 cpufreq_set_update_util_data(i, NULL);
196 static void gov_cancel_work(struct policy_dbs_info *policy_dbs)
198 /* Tell dbs_update_util_handler() to skip queuing up work items. */
199 atomic_inc(&policy_dbs->work_count);
201 * If dbs_update_util_handler() is already running, it may not notice
202 * the incremented work_count, so wait for it to complete to prevent its
203 * work item from being queued up after the cancel_work_sync() below.
205 gov_clear_update_util(policy_dbs->policy);
206 irq_work_sync(&policy_dbs->irq_work);
207 cancel_work_sync(&policy_dbs->work);
208 atomic_set(&policy_dbs->work_count, 0);
211 static void dbs_work_handler(struct work_struct *work)
213 struct policy_dbs_info *policy_dbs;
214 struct cpufreq_policy *policy;
215 struct dbs_governor *gov;
218 policy_dbs = container_of(work, struct policy_dbs_info, work);
219 policy = policy_dbs->policy;
220 gov = dbs_governor_of(policy);
223 * Make sure cpufreq_governor_limits() isn't evaluating load or the
224 * ondemand governor isn't updating the sampling rate in parallel.
226 mutex_lock(&policy_dbs->timer_mutex);
227 delay = gov->gov_dbs_timer(policy);
228 policy_dbs->sample_delay_ns = jiffies_to_nsecs(delay);
229 mutex_unlock(&policy_dbs->timer_mutex);
232 * If the atomic operation below is reordered with respect to the
233 * sample delay modification, the utilization update handler may end
234 * up using a stale sample delay value.
236 smp_mb__before_atomic();
237 atomic_dec(&policy_dbs->work_count);
240 static void dbs_irq_work(struct irq_work *irq_work)
242 struct policy_dbs_info *policy_dbs;
244 policy_dbs = container_of(irq_work, struct policy_dbs_info, irq_work);
245 schedule_work(&policy_dbs->work);
248 static inline void gov_queue_irq_work(struct policy_dbs_info *policy_dbs)
251 irq_work_queue_on(&policy_dbs->irq_work, smp_processor_id());
253 irq_work_queue(&policy_dbs->irq_work);
257 static void dbs_update_util_handler(struct update_util_data *data, u64 time,
258 unsigned long util, unsigned long max)
260 struct cpu_dbs_info *cdbs = container_of(data, struct cpu_dbs_info, update_util);
261 struct policy_dbs_info *policy_dbs = cdbs->policy_dbs;
264 * The work may not be allowed to be queued up right now.
266 * - Work has already been queued up or is in progress.
267 * - The governor is being stopped.
268 * - It is too early (too little time from the previous sample).
270 if (atomic_inc_return(&policy_dbs->work_count) == 1) {
273 delta_ns = time - policy_dbs->last_sample_time;
274 if ((s64)delta_ns >= policy_dbs->sample_delay_ns) {
275 policy_dbs->last_sample_time = time;
276 gov_queue_irq_work(policy_dbs);
280 atomic_dec(&policy_dbs->work_count);
283 static void set_sampling_rate(struct dbs_data *dbs_data,
284 struct dbs_governor *gov,
285 unsigned int sampling_rate)
287 if (gov->governor == GOV_CONSERVATIVE) {
288 struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
289 cs_tuners->sampling_rate = sampling_rate;
291 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
292 od_tuners->sampling_rate = sampling_rate;
296 static struct policy_dbs_info *alloc_policy_dbs_info(struct cpufreq_policy *policy,
297 struct dbs_governor *gov)
299 struct policy_dbs_info *policy_dbs;
302 /* Allocate memory for the common information for policy->cpus */
303 policy_dbs = kzalloc(sizeof(*policy_dbs), GFP_KERNEL);
307 mutex_init(&policy_dbs->timer_mutex);
308 atomic_set(&policy_dbs->work_count, 0);
309 init_irq_work(&policy_dbs->irq_work, dbs_irq_work);
310 INIT_WORK(&policy_dbs->work, dbs_work_handler);
312 /* Set policy_dbs for all CPUs, online+offline */
313 for_each_cpu(j, policy->related_cpus) {
314 struct cpu_dbs_info *j_cdbs = gov->get_cpu_cdbs(j);
316 j_cdbs->policy_dbs = policy_dbs;
317 j_cdbs->update_util.func = dbs_update_util_handler;
322 static void free_policy_dbs_info(struct cpufreq_policy *policy,
323 struct dbs_governor *gov)
325 struct cpu_dbs_info *cdbs = gov->get_cpu_cdbs(policy->cpu);
326 struct policy_dbs_info *policy_dbs = cdbs->policy_dbs;
329 mutex_destroy(&policy_dbs->timer_mutex);
331 for_each_cpu(j, policy->related_cpus) {
332 struct cpu_dbs_info *j_cdbs = gov->get_cpu_cdbs(j);
334 j_cdbs->policy_dbs = NULL;
335 j_cdbs->update_util.func = NULL;
340 static int cpufreq_governor_init(struct cpufreq_policy *policy)
342 struct dbs_governor *gov = dbs_governor_of(policy);
343 struct dbs_data *dbs_data = gov->gdbs_data;
344 struct policy_dbs_info *policy_dbs;
345 unsigned int latency;
348 /* State should be equivalent to EXIT */
349 if (policy->governor_data)
352 policy_dbs = alloc_policy_dbs_info(policy, gov);
357 if (WARN_ON(have_governor_per_policy())) {
359 goto free_policy_dbs_info;
361 dbs_data->usage_count++;
362 policy_dbs->dbs_data = dbs_data;
363 policy->governor_data = policy_dbs;
367 dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
370 goto free_policy_dbs_info;
373 dbs_data->usage_count = 1;
375 ret = gov->init(dbs_data, !policy->governor->initialized);
377 goto free_policy_dbs_info;
379 /* policy latency is in ns. Convert it to us first */
380 latency = policy->cpuinfo.transition_latency / 1000;
384 /* Bring kernel and HW constraints together */
385 dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate,
386 MIN_LATENCY_MULTIPLIER * latency);
387 set_sampling_rate(dbs_data, gov, max(dbs_data->min_sampling_rate,
388 latency * LATENCY_MULTIPLIER));
390 if (!have_governor_per_policy())
391 gov->gdbs_data = dbs_data;
393 policy_dbs->dbs_data = dbs_data;
394 policy->governor_data = policy_dbs;
396 ret = sysfs_create_group(get_governor_parent_kobj(policy),
397 get_sysfs_attr(gov));
401 /* Failure, so roll back. */
403 policy->governor_data = NULL;
405 if (!have_governor_per_policy())
406 gov->gdbs_data = NULL;
407 gov->exit(dbs_data, !policy->governor->initialized);
410 free_policy_dbs_info:
411 free_policy_dbs_info(policy, gov);
415 static int cpufreq_governor_exit(struct cpufreq_policy *policy)
417 struct dbs_governor *gov = dbs_governor_of(policy);
418 struct policy_dbs_info *policy_dbs = policy->governor_data;
419 struct dbs_data *dbs_data = policy_dbs->dbs_data;
421 /* State should be equivalent to INIT */
422 if (policy_dbs->policy)
425 if (!--dbs_data->usage_count) {
426 sysfs_remove_group(get_governor_parent_kobj(policy),
427 get_sysfs_attr(gov));
429 policy->governor_data = NULL;
431 if (!have_governor_per_policy())
432 gov->gdbs_data = NULL;
434 gov->exit(dbs_data, policy->governor->initialized == 1);
437 policy->governor_data = NULL;
440 free_policy_dbs_info(policy, gov);
444 static int cpufreq_governor_start(struct cpufreq_policy *policy)
446 struct dbs_governor *gov = dbs_governor_of(policy);
447 struct policy_dbs_info *policy_dbs = policy->governor_data;
448 struct dbs_data *dbs_data = policy_dbs->dbs_data;
449 unsigned int sampling_rate, ignore_nice, j, cpu = policy->cpu;
455 /* State should be equivalent to INIT */
456 if (policy_dbs->policy)
459 if (gov->governor == GOV_CONSERVATIVE) {
460 struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
462 sampling_rate = cs_tuners->sampling_rate;
463 ignore_nice = cs_tuners->ignore_nice_load;
465 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
467 sampling_rate = od_tuners->sampling_rate;
468 ignore_nice = od_tuners->ignore_nice_load;
469 io_busy = od_tuners->io_is_busy;
472 for_each_cpu(j, policy->cpus) {
473 struct cpu_dbs_info *j_cdbs = gov->get_cpu_cdbs(j);
474 unsigned int prev_load;
476 j_cdbs->prev_cpu_idle =
477 get_cpu_idle_time(j, &j_cdbs->prev_cpu_wall, io_busy);
479 prev_load = (unsigned int)(j_cdbs->prev_cpu_wall -
480 j_cdbs->prev_cpu_idle);
481 j_cdbs->prev_load = 100 * prev_load /
482 (unsigned int)j_cdbs->prev_cpu_wall;
485 j_cdbs->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
487 policy_dbs->policy = policy;
489 if (gov->governor == GOV_CONSERVATIVE) {
490 struct cs_cpu_dbs_info_s *cs_dbs_info =
491 gov->get_cpu_dbs_info_s(cpu);
493 cs_dbs_info->down_skip = 0;
494 cs_dbs_info->requested_freq = policy->cur;
496 struct od_ops *od_ops = gov->gov_ops;
497 struct od_cpu_dbs_info_s *od_dbs_info = gov->get_cpu_dbs_info_s(cpu);
499 od_dbs_info->rate_mult = 1;
500 od_dbs_info->sample_type = OD_NORMAL_SAMPLE;
501 od_ops->powersave_bias_init_cpu(cpu);
504 gov_set_update_util(policy_dbs, sampling_rate);
508 static int cpufreq_governor_stop(struct cpufreq_policy *policy)
510 struct policy_dbs_info *policy_dbs = policy->governor_data;
512 /* State should be equivalent to START */
513 if (!policy_dbs->policy)
516 gov_cancel_work(policy_dbs);
517 policy_dbs->policy = NULL;
522 static int cpufreq_governor_limits(struct cpufreq_policy *policy)
524 struct policy_dbs_info *policy_dbs = policy->governor_data;
526 /* State should be equivalent to START */
527 if (!policy_dbs->policy)
530 mutex_lock(&policy_dbs->timer_mutex);
531 if (policy->max < policy->cur)
532 __cpufreq_driver_target(policy, policy->max, CPUFREQ_RELATION_H);
533 else if (policy->min > policy->cur)
534 __cpufreq_driver_target(policy, policy->min, CPUFREQ_RELATION_L);
535 dbs_check_cpu(policy);
536 mutex_unlock(&policy_dbs->timer_mutex);
541 int cpufreq_governor_dbs(struct cpufreq_policy *policy, unsigned int event)
545 /* Lock governor to block concurrent initialization of governor */
546 mutex_lock(&dbs_data_mutex);
548 if (event == CPUFREQ_GOV_POLICY_INIT) {
549 ret = cpufreq_governor_init(policy);
550 } else if (policy->governor_data) {
552 case CPUFREQ_GOV_POLICY_EXIT:
553 ret = cpufreq_governor_exit(policy);
555 case CPUFREQ_GOV_START:
556 ret = cpufreq_governor_start(policy);
558 case CPUFREQ_GOV_STOP:
559 ret = cpufreq_governor_stop(policy);
561 case CPUFREQ_GOV_LIMITS:
562 ret = cpufreq_governor_limits(policy);
567 mutex_unlock(&dbs_data_mutex);
570 EXPORT_SYMBOL_GPL(cpufreq_governor_dbs);