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 static struct attribute_group *get_sysfs_attr(struct dbs_data *dbs_data)
27 if (have_governor_per_policy())
28 return dbs_data->cdata->attr_group_gov_pol;
30 return dbs_data->cdata->attr_group_gov_sys;
33 void dbs_check_cpu(struct dbs_data *dbs_data, int cpu)
35 struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
36 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
37 struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
38 struct cpufreq_policy *policy;
39 unsigned int sampling_rate;
40 unsigned int max_load = 0;
41 unsigned int ignore_nice;
44 if (dbs_data->cdata->governor == GOV_ONDEMAND) {
45 struct od_cpu_dbs_info_s *od_dbs_info =
46 dbs_data->cdata->get_cpu_dbs_info_s(cpu);
49 * Sometimes, the ondemand governor uses an additional
50 * multiplier to give long delays. So apply this multiplier to
51 * the 'sampling_rate', so as to keep the wake-up-from-idle
52 * detection logic a bit conservative.
54 sampling_rate = od_tuners->sampling_rate;
55 sampling_rate *= od_dbs_info->rate_mult;
57 ignore_nice = od_tuners->ignore_nice_load;
59 sampling_rate = cs_tuners->sampling_rate;
60 ignore_nice = cs_tuners->ignore_nice_load;
63 policy = cdbs->cur_policy;
65 /* Get Absolute Load */
66 for_each_cpu(j, policy->cpus) {
67 struct cpu_dbs_common_info *j_cdbs;
68 u64 cur_wall_time, cur_idle_time;
69 unsigned int idle_time, wall_time;
73 j_cdbs = dbs_data->cdata->get_cpu_cdbs(j);
76 * For the purpose of ondemand, waiting for disk IO is
77 * an indication that you're performance critical, and
78 * not that the system is actually idle. So do not add
79 * the iowait time to the cpu idle time.
81 if (dbs_data->cdata->governor == GOV_ONDEMAND)
82 io_busy = od_tuners->io_is_busy;
83 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time, io_busy);
85 wall_time = (unsigned int)
86 (cur_wall_time - j_cdbs->prev_cpu_wall);
87 j_cdbs->prev_cpu_wall = cur_wall_time;
89 idle_time = (unsigned int)
90 (cur_idle_time - j_cdbs->prev_cpu_idle);
91 j_cdbs->prev_cpu_idle = cur_idle_time;
95 unsigned long cur_nice_jiffies;
97 cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] -
100 * Assumption: nice time between sampling periods will
101 * be less than 2^32 jiffies for 32 bit sys
103 cur_nice_jiffies = (unsigned long)
104 cputime64_to_jiffies64(cur_nice);
106 cdbs->prev_cpu_nice =
107 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
108 idle_time += jiffies_to_usecs(cur_nice_jiffies);
111 if (unlikely(!wall_time || wall_time < idle_time))
115 * If the CPU had gone completely idle, and a task just woke up
116 * on this CPU now, it would be unfair to calculate 'load' the
117 * usual way for this elapsed time-window, because it will show
118 * near-zero load, irrespective of how CPU intensive that task
119 * actually is. This is undesirable for latency-sensitive bursty
122 * To avoid this, we reuse the 'load' from the previous
123 * time-window and give this task a chance to start with a
124 * reasonably high CPU frequency. (However, we shouldn't over-do
125 * this copy, lest we get stuck at a high load (high frequency)
126 * for too long, even when the current system load has actually
127 * dropped down. So we perform the copy only once, upon the
128 * first wake-up from idle.)
130 * Detecting this situation is easy: the governor's deferrable
131 * timer would not have fired during CPU-idle periods. Hence
132 * an unusually large 'wall_time' (as compared to the sampling
133 * rate) indicates this scenario.
135 if (unlikely(wall_time > (2 * sampling_rate)) &&
136 j_cdbs->copy_prev_load) {
137 load = j_cdbs->prev_load;
138 j_cdbs->copy_prev_load = false;
140 load = 100 * (wall_time - idle_time) / wall_time;
141 j_cdbs->prev_load = load;
142 j_cdbs->copy_prev_load = true;
149 dbs_data->cdata->gov_check_cpu(cpu, max_load);
151 EXPORT_SYMBOL_GPL(dbs_check_cpu);
153 static inline void __gov_queue_work(int cpu, struct dbs_data *dbs_data,
156 struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
158 mod_delayed_work_on(cpu, system_wq, &cdbs->work, delay);
161 void gov_queue_work(struct dbs_data *dbs_data, struct cpufreq_policy *policy,
162 unsigned int delay, bool all_cpus)
166 mutex_lock(&cpufreq_governor_lock);
167 if (!policy->governor_enabled)
172 * Use raw_smp_processor_id() to avoid preemptible warnings.
173 * We know that this is only called with all_cpus == false from
174 * works that have been queued with *_work_on() functions and
175 * those works are canceled during CPU_DOWN_PREPARE so they
176 * can't possibly run on any other CPU.
178 __gov_queue_work(raw_smp_processor_id(), dbs_data, delay);
180 for_each_cpu(i, policy->cpus)
181 __gov_queue_work(i, dbs_data, delay);
185 mutex_unlock(&cpufreq_governor_lock);
187 EXPORT_SYMBOL_GPL(gov_queue_work);
189 static inline void gov_cancel_work(struct dbs_data *dbs_data,
190 struct cpufreq_policy *policy)
192 struct cpu_dbs_common_info *cdbs;
195 for_each_cpu(i, policy->cpus) {
196 cdbs = dbs_data->cdata->get_cpu_cdbs(i);
197 cancel_delayed_work_sync(&cdbs->work);
201 /* Will return if we need to evaluate cpu load again or not */
202 bool need_load_eval(struct cpu_dbs_common_info *cdbs,
203 unsigned int sampling_rate)
205 if (policy_is_shared(cdbs->cur_policy)) {
206 ktime_t time_now = ktime_get();
207 s64 delta_us = ktime_us_delta(time_now, cdbs->time_stamp);
209 /* Do nothing if we recently have sampled */
210 if (delta_us < (s64)(sampling_rate / 2))
213 cdbs->time_stamp = time_now;
218 EXPORT_SYMBOL_GPL(need_load_eval);
220 static void set_sampling_rate(struct dbs_data *dbs_data,
221 unsigned int sampling_rate)
223 if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
224 struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
225 cs_tuners->sampling_rate = sampling_rate;
227 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
228 od_tuners->sampling_rate = sampling_rate;
232 int cpufreq_governor_dbs(struct cpufreq_policy *policy,
233 struct common_dbs_data *cdata, unsigned int event)
235 struct dbs_data *dbs_data;
236 struct od_cpu_dbs_info_s *od_dbs_info = NULL;
237 struct cs_cpu_dbs_info_s *cs_dbs_info = NULL;
238 struct od_ops *od_ops = NULL;
239 struct od_dbs_tuners *od_tuners = NULL;
240 struct cs_dbs_tuners *cs_tuners = NULL;
241 struct cpu_dbs_common_info *cpu_cdbs;
242 unsigned int sampling_rate, latency, ignore_nice, j, cpu = policy->cpu;
246 if (have_governor_per_policy())
247 dbs_data = policy->governor_data;
249 dbs_data = cdata->gdbs_data;
251 WARN_ON(!dbs_data && (event != CPUFREQ_GOV_POLICY_INIT));
254 case CPUFREQ_GOV_POLICY_INIT:
255 if (have_governor_per_policy()) {
257 } else if (dbs_data) {
258 dbs_data->usage_count++;
259 policy->governor_data = dbs_data;
263 dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
265 pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
269 dbs_data->cdata = cdata;
270 dbs_data->usage_count = 1;
271 rc = cdata->init(dbs_data);
273 pr_err("%s: POLICY_INIT: init() failed\n", __func__);
278 if (!have_governor_per_policy())
279 WARN_ON(cpufreq_get_global_kobject());
281 rc = sysfs_create_group(get_governor_parent_kobj(policy),
282 get_sysfs_attr(dbs_data));
284 cdata->exit(dbs_data);
289 policy->governor_data = dbs_data;
291 /* policy latency is in ns. Convert it to us first */
292 latency = policy->cpuinfo.transition_latency / 1000;
296 /* Bring kernel and HW constraints together */
297 dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate,
298 MIN_LATENCY_MULTIPLIER * latency);
299 set_sampling_rate(dbs_data, max(dbs_data->min_sampling_rate,
300 latency * LATENCY_MULTIPLIER));
302 if ((cdata->governor == GOV_CONSERVATIVE) &&
303 (!policy->governor->initialized)) {
304 struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
306 cpufreq_register_notifier(cs_ops->notifier_block,
307 CPUFREQ_TRANSITION_NOTIFIER);
310 if (!have_governor_per_policy())
311 cdata->gdbs_data = dbs_data;
314 case CPUFREQ_GOV_POLICY_EXIT:
315 if (!--dbs_data->usage_count) {
316 sysfs_remove_group(get_governor_parent_kobj(policy),
317 get_sysfs_attr(dbs_data));
319 if (!have_governor_per_policy())
320 cpufreq_put_global_kobject();
322 if ((dbs_data->cdata->governor == GOV_CONSERVATIVE) &&
323 (policy->governor->initialized == 1)) {
324 struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
326 cpufreq_unregister_notifier(cs_ops->notifier_block,
327 CPUFREQ_TRANSITION_NOTIFIER);
330 cdata->exit(dbs_data);
332 cdata->gdbs_data = NULL;
335 policy->governor_data = NULL;
339 cpu_cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
341 if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
342 cs_tuners = dbs_data->tuners;
343 cs_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
344 sampling_rate = cs_tuners->sampling_rate;
345 ignore_nice = cs_tuners->ignore_nice_load;
347 od_tuners = dbs_data->tuners;
348 od_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
349 sampling_rate = od_tuners->sampling_rate;
350 ignore_nice = od_tuners->ignore_nice_load;
351 od_ops = dbs_data->cdata->gov_ops;
352 io_busy = od_tuners->io_is_busy;
356 case CPUFREQ_GOV_START:
360 mutex_lock(&dbs_data->mutex);
362 for_each_cpu(j, policy->cpus) {
363 struct cpu_dbs_common_info *j_cdbs =
364 dbs_data->cdata->get_cpu_cdbs(j);
365 unsigned int prev_load;
368 j_cdbs->cur_policy = policy;
369 j_cdbs->prev_cpu_idle = get_cpu_idle_time(j,
370 &j_cdbs->prev_cpu_wall, io_busy);
372 prev_load = (unsigned int)
373 (j_cdbs->prev_cpu_wall - j_cdbs->prev_cpu_idle);
374 j_cdbs->prev_load = 100 * prev_load /
375 (unsigned int) j_cdbs->prev_cpu_wall;
376 j_cdbs->copy_prev_load = true;
379 j_cdbs->prev_cpu_nice =
380 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
382 mutex_init(&j_cdbs->timer_mutex);
383 INIT_DEFERRABLE_WORK(&j_cdbs->work,
384 dbs_data->cdata->gov_dbs_timer);
387 if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
388 cs_dbs_info->down_skip = 0;
389 cs_dbs_info->enable = 1;
390 cs_dbs_info->requested_freq = policy->cur;
392 od_dbs_info->rate_mult = 1;
393 od_dbs_info->sample_type = OD_NORMAL_SAMPLE;
394 od_ops->powersave_bias_init_cpu(cpu);
397 mutex_unlock(&dbs_data->mutex);
399 /* Initiate timer time stamp */
400 cpu_cdbs->time_stamp = ktime_get();
402 gov_queue_work(dbs_data, policy,
403 delay_for_sampling_rate(sampling_rate), true);
406 case CPUFREQ_GOV_STOP:
407 if (dbs_data->cdata->governor == GOV_CONSERVATIVE)
408 cs_dbs_info->enable = 0;
410 gov_cancel_work(dbs_data, policy);
412 mutex_lock(&dbs_data->mutex);
413 mutex_destroy(&cpu_cdbs->timer_mutex);
414 cpu_cdbs->cur_policy = NULL;
416 mutex_unlock(&dbs_data->mutex);
420 case CPUFREQ_GOV_LIMITS:
421 mutex_lock(&dbs_data->mutex);
422 if (!cpu_cdbs->cur_policy) {
423 mutex_unlock(&dbs_data->mutex);
426 mutex_lock(&cpu_cdbs->timer_mutex);
427 if (policy->max < cpu_cdbs->cur_policy->cur)
428 __cpufreq_driver_target(cpu_cdbs->cur_policy,
429 policy->max, CPUFREQ_RELATION_H);
430 else if (policy->min > cpu_cdbs->cur_policy->cur)
431 __cpufreq_driver_target(cpu_cdbs->cur_policy,
432 policy->min, CPUFREQ_RELATION_L);
433 dbs_check_cpu(dbs_data, cpu);
434 mutex_unlock(&cpu_cdbs->timer_mutex);
435 mutex_unlock(&dbs_data->mutex);
440 EXPORT_SYMBOL_GPL(cpufreq_governor_dbs);