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 <asm/cputime.h>
20 #include <linux/cpufreq.h>
21 #include <linux/cpumask.h>
22 #include <linux/export.h>
23 #include <linux/kernel_stat.h>
24 #include <linux/mutex.h>
25 #include <linux/slab.h>
26 #include <linux/types.h>
27 #include <linux/workqueue.h>
28 #include <linux/cpu.h>
30 #include "cpufreq_governor.h"
32 static struct attribute_group *get_sysfs_attr(struct dbs_data *dbs_data)
34 if (have_governor_per_policy())
35 return dbs_data->cdata->attr_group_gov_pol;
37 return dbs_data->cdata->attr_group_gov_sys;
40 void dbs_check_cpu(struct dbs_data *dbs_data, int cpu)
42 struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
43 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
44 struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
45 struct cpufreq_policy *policy;
46 unsigned int max_load = 0;
47 unsigned int ignore_nice;
50 if (dbs_data->cdata->governor == GOV_ONDEMAND)
51 ignore_nice = od_tuners->ignore_nice;
53 ignore_nice = cs_tuners->ignore_nice;
55 policy = cdbs->cur_policy;
57 /* Get Absolute Load (in terms of freq for ondemand gov) */
58 for_each_cpu(j, policy->cpus) {
59 struct cpu_dbs_common_info *j_cdbs;
60 u64 cur_wall_time, cur_idle_time;
61 unsigned int idle_time, wall_time;
65 j_cdbs = dbs_data->cdata->get_cpu_cdbs(j);
68 * For the purpose of ondemand, waiting for disk IO is
69 * an indication that you're performance critical, and
70 * not that the system is actually idle. So do not add
71 * the iowait time to the cpu idle time.
73 if (dbs_data->cdata->governor == GOV_ONDEMAND)
74 io_busy = od_tuners->io_is_busy;
75 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time, io_busy);
77 wall_time = (unsigned int)
78 (cur_wall_time - j_cdbs->prev_cpu_wall);
79 j_cdbs->prev_cpu_wall = cur_wall_time;
81 idle_time = (unsigned int)
82 (cur_idle_time - j_cdbs->prev_cpu_idle);
83 j_cdbs->prev_cpu_idle = cur_idle_time;
87 unsigned long cur_nice_jiffies;
89 cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] -
92 * Assumption: nice time between sampling periods will
93 * be less than 2^32 jiffies for 32 bit sys
95 cur_nice_jiffies = (unsigned long)
96 cputime64_to_jiffies64(cur_nice);
99 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
100 idle_time += jiffies_to_usecs(cur_nice_jiffies);
103 if (unlikely(!wall_time || wall_time < idle_time))
106 load = 100 * (wall_time - idle_time) / wall_time;
108 if (dbs_data->cdata->governor == GOV_ONDEMAND) {
109 int freq_avg = __cpufreq_driver_getavg(policy, j);
111 freq_avg = policy->cur;
120 dbs_data->cdata->gov_check_cpu(cpu, max_load);
122 EXPORT_SYMBOL_GPL(dbs_check_cpu);
124 static inline void __gov_queue_work(int cpu, struct dbs_data *dbs_data,
127 struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
129 mod_delayed_work_on(cpu, system_wq, &cdbs->work, delay);
132 void gov_queue_work(struct dbs_data *dbs_data, struct cpufreq_policy *policy,
133 unsigned int delay, bool all_cpus)
138 __gov_queue_work(smp_processor_id(), dbs_data, delay);
141 for_each_cpu(i, policy->cpus)
142 __gov_queue_work(i, dbs_data, delay);
146 EXPORT_SYMBOL_GPL(gov_queue_work);
148 static inline void gov_cancel_work(struct dbs_data *dbs_data,
149 struct cpufreq_policy *policy)
151 struct cpu_dbs_common_info *cdbs;
154 for_each_cpu(i, policy->cpus) {
155 cdbs = dbs_data->cdata->get_cpu_cdbs(i);
156 cancel_delayed_work_sync(&cdbs->work);
160 /* Will return if we need to evaluate cpu load again or not */
161 bool need_load_eval(struct cpu_dbs_common_info *cdbs,
162 unsigned int sampling_rate)
164 if (policy_is_shared(cdbs->cur_policy)) {
165 ktime_t time_now = ktime_get();
166 s64 delta_us = ktime_us_delta(time_now, cdbs->time_stamp);
168 /* Do nothing if we recently have sampled */
169 if (delta_us < (s64)(sampling_rate / 2))
172 cdbs->time_stamp = time_now;
177 EXPORT_SYMBOL_GPL(need_load_eval);
179 static void set_sampling_rate(struct dbs_data *dbs_data,
180 unsigned int sampling_rate)
182 if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
183 struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
184 cs_tuners->sampling_rate = sampling_rate;
186 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
187 od_tuners->sampling_rate = sampling_rate;
191 int cpufreq_governor_dbs(struct cpufreq_policy *policy,
192 struct common_dbs_data *cdata, unsigned int event)
194 struct dbs_data *dbs_data;
195 struct od_cpu_dbs_info_s *od_dbs_info = NULL;
196 struct cs_cpu_dbs_info_s *cs_dbs_info = NULL;
197 struct od_ops *od_ops = NULL;
198 struct od_dbs_tuners *od_tuners = NULL;
199 struct cs_dbs_tuners *cs_tuners = NULL;
200 struct cpu_dbs_common_info *cpu_cdbs;
201 unsigned int sampling_rate, latency, ignore_nice, j, cpu = policy->cpu;
205 if (have_governor_per_policy())
206 dbs_data = policy->governor_data;
208 dbs_data = cdata->gdbs_data;
210 WARN_ON(!dbs_data && (event != CPUFREQ_GOV_POLICY_INIT));
213 case CPUFREQ_GOV_POLICY_INIT:
214 if (have_governor_per_policy()) {
216 } else if (dbs_data) {
217 dbs_data->usage_count++;
218 policy->governor_data = dbs_data;
222 dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
224 pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
228 dbs_data->cdata = cdata;
229 dbs_data->usage_count = 1;
230 rc = cdata->init(dbs_data);
232 pr_err("%s: POLICY_INIT: init() failed\n", __func__);
237 rc = sysfs_create_group(get_governor_parent_kobj(policy),
238 get_sysfs_attr(dbs_data));
240 cdata->exit(dbs_data);
245 policy->governor_data = dbs_data;
247 /* policy latency is in nS. Convert it to uS first */
248 latency = policy->cpuinfo.transition_latency / 1000;
252 /* Bring kernel and HW constraints together */
253 dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate,
254 MIN_LATENCY_MULTIPLIER * latency);
255 set_sampling_rate(dbs_data, max(dbs_data->min_sampling_rate,
256 latency * LATENCY_MULTIPLIER));
258 if ((cdata->governor == GOV_CONSERVATIVE) &&
259 (!policy->governor->initialized)) {
260 struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
262 cpufreq_register_notifier(cs_ops->notifier_block,
263 CPUFREQ_TRANSITION_NOTIFIER);
266 if (!have_governor_per_policy())
267 cdata->gdbs_data = dbs_data;
270 case CPUFREQ_GOV_POLICY_EXIT:
271 if (!--dbs_data->usage_count) {
272 sysfs_remove_group(get_governor_parent_kobj(policy),
273 get_sysfs_attr(dbs_data));
275 if ((dbs_data->cdata->governor == GOV_CONSERVATIVE) &&
276 (policy->governor->initialized == 1)) {
277 struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
279 cpufreq_unregister_notifier(cs_ops->notifier_block,
280 CPUFREQ_TRANSITION_NOTIFIER);
283 cdata->exit(dbs_data);
285 cdata->gdbs_data = NULL;
288 policy->governor_data = NULL;
292 cpu_cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
294 if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
295 cs_tuners = dbs_data->tuners;
296 cs_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
297 sampling_rate = cs_tuners->sampling_rate;
298 ignore_nice = cs_tuners->ignore_nice;
300 od_tuners = dbs_data->tuners;
301 od_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
302 sampling_rate = od_tuners->sampling_rate;
303 ignore_nice = od_tuners->ignore_nice;
304 od_ops = dbs_data->cdata->gov_ops;
305 io_busy = od_tuners->io_is_busy;
309 case CPUFREQ_GOV_START:
313 mutex_lock(&dbs_data->mutex);
315 for_each_cpu(j, policy->cpus) {
316 struct cpu_dbs_common_info *j_cdbs =
317 dbs_data->cdata->get_cpu_cdbs(j);
320 j_cdbs->cur_policy = policy;
321 j_cdbs->prev_cpu_idle = get_cpu_idle_time(j,
322 &j_cdbs->prev_cpu_wall, io_busy);
324 j_cdbs->prev_cpu_nice =
325 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
327 mutex_init(&j_cdbs->timer_mutex);
328 INIT_DEFERRABLE_WORK(&j_cdbs->work,
329 dbs_data->cdata->gov_dbs_timer);
333 * conservative does not implement micro like ondemand
334 * governor, thus we are bound to jiffes/HZ
336 if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
337 cs_dbs_info->down_skip = 0;
338 cs_dbs_info->enable = 1;
339 cs_dbs_info->requested_freq = policy->cur;
341 od_dbs_info->rate_mult = 1;
342 od_dbs_info->sample_type = OD_NORMAL_SAMPLE;
343 od_ops->powersave_bias_init_cpu(cpu);
346 mutex_unlock(&dbs_data->mutex);
348 /* Initiate timer time stamp */
349 cpu_cdbs->time_stamp = ktime_get();
351 gov_queue_work(dbs_data, policy,
352 delay_for_sampling_rate(sampling_rate), true);
355 case CPUFREQ_GOV_STOP:
356 if (dbs_data->cdata->governor == GOV_CONSERVATIVE)
357 cs_dbs_info->enable = 0;
359 gov_cancel_work(dbs_data, policy);
361 mutex_lock(&dbs_data->mutex);
362 mutex_destroy(&cpu_cdbs->timer_mutex);
364 mutex_unlock(&dbs_data->mutex);
368 case CPUFREQ_GOV_LIMITS:
369 mutex_lock(&cpu_cdbs->timer_mutex);
370 if (policy->max < cpu_cdbs->cur_policy->cur)
371 __cpufreq_driver_target(cpu_cdbs->cur_policy,
372 policy->max, CPUFREQ_RELATION_H);
373 else if (policy->min > cpu_cdbs->cur_policy->cur)
374 __cpufreq_driver_target(cpu_cdbs->cur_policy,
375 policy->min, CPUFREQ_RELATION_L);
376 dbs_check_cpu(dbs_data, cpu);
377 mutex_unlock(&cpu_cdbs->timer_mutex);
382 EXPORT_SYMBOL_GPL(cpufreq_governor_dbs);