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 max_load = 0;
40 unsigned int ignore_nice;
43 if (dbs_data->cdata->governor == GOV_ONDEMAND)
44 ignore_nice = od_tuners->ignore_nice_load;
46 ignore_nice = cs_tuners->ignore_nice_load;
48 policy = cdbs->cur_policy;
50 /* Get Absolute Load */
51 for_each_cpu(j, policy->cpus) {
52 struct cpu_dbs_common_info *j_cdbs;
53 u64 cur_wall_time, cur_idle_time;
54 unsigned int idle_time, wall_time;
58 j_cdbs = dbs_data->cdata->get_cpu_cdbs(j);
61 * For the purpose of ondemand, waiting for disk IO is
62 * an indication that you're performance critical, and
63 * not that the system is actually idle. So do not add
64 * the iowait time to the cpu idle time.
66 if (dbs_data->cdata->governor == GOV_ONDEMAND)
67 io_busy = od_tuners->io_is_busy;
68 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time, io_busy);
70 wall_time = (unsigned int)
71 (cur_wall_time - j_cdbs->prev_cpu_wall);
72 j_cdbs->prev_cpu_wall = cur_wall_time;
74 idle_time = (unsigned int)
75 (cur_idle_time - j_cdbs->prev_cpu_idle);
76 j_cdbs->prev_cpu_idle = cur_idle_time;
80 unsigned long cur_nice_jiffies;
82 cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] -
85 * Assumption: nice time between sampling periods will
86 * be less than 2^32 jiffies for 32 bit sys
88 cur_nice_jiffies = (unsigned long)
89 cputime64_to_jiffies64(cur_nice);
92 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
93 idle_time += jiffies_to_usecs(cur_nice_jiffies);
96 if (unlikely(!wall_time || wall_time < idle_time))
99 load = 100 * (wall_time - idle_time) / wall_time;
105 dbs_data->cdata->gov_check_cpu(cpu, max_load);
107 EXPORT_SYMBOL_GPL(dbs_check_cpu);
109 static inline void __gov_queue_work(int cpu, struct dbs_data *dbs_data,
112 struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
114 mod_delayed_work_on(cpu, system_wq, &cdbs->work, delay);
117 void gov_queue_work(struct dbs_data *dbs_data, struct cpufreq_policy *policy,
118 unsigned int delay, bool all_cpus)
122 if (!policy->governor_enabled)
126 __gov_queue_work(smp_processor_id(), dbs_data, delay);
128 for_each_cpu(i, policy->cpus)
129 __gov_queue_work(i, dbs_data, delay);
132 EXPORT_SYMBOL_GPL(gov_queue_work);
134 static inline void gov_cancel_work(struct dbs_data *dbs_data,
135 struct cpufreq_policy *policy)
137 struct cpu_dbs_common_info *cdbs;
140 for_each_cpu(i, policy->cpus) {
141 cdbs = dbs_data->cdata->get_cpu_cdbs(i);
142 cancel_delayed_work_sync(&cdbs->work);
146 /* Will return if we need to evaluate cpu load again or not */
147 bool need_load_eval(struct cpu_dbs_common_info *cdbs,
148 unsigned int sampling_rate)
150 if (policy_is_shared(cdbs->cur_policy)) {
151 ktime_t time_now = ktime_get();
152 s64 delta_us = ktime_us_delta(time_now, cdbs->time_stamp);
154 /* Do nothing if we recently have sampled */
155 if (delta_us < (s64)(sampling_rate / 2))
158 cdbs->time_stamp = time_now;
163 EXPORT_SYMBOL_GPL(need_load_eval);
165 static void set_sampling_rate(struct dbs_data *dbs_data,
166 unsigned int sampling_rate)
168 if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
169 struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
170 cs_tuners->sampling_rate = sampling_rate;
172 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
173 od_tuners->sampling_rate = sampling_rate;
177 int cpufreq_governor_dbs(struct cpufreq_policy *policy,
178 struct common_dbs_data *cdata, unsigned int event)
180 struct dbs_data *dbs_data;
181 struct od_cpu_dbs_info_s *od_dbs_info = NULL;
182 struct cs_cpu_dbs_info_s *cs_dbs_info = NULL;
183 struct od_ops *od_ops = NULL;
184 struct od_dbs_tuners *od_tuners = NULL;
185 struct cs_dbs_tuners *cs_tuners = NULL;
186 struct cpu_dbs_common_info *cpu_cdbs;
187 unsigned int sampling_rate, latency, ignore_nice, j, cpu = policy->cpu;
191 if (have_governor_per_policy())
192 dbs_data = policy->governor_data;
194 dbs_data = cdata->gdbs_data;
196 WARN_ON(!dbs_data && (event != CPUFREQ_GOV_POLICY_INIT));
199 case CPUFREQ_GOV_POLICY_INIT:
200 if (have_governor_per_policy()) {
202 } else if (dbs_data) {
203 dbs_data->usage_count++;
204 policy->governor_data = dbs_data;
208 dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
210 pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
214 dbs_data->cdata = cdata;
215 dbs_data->usage_count = 1;
216 rc = cdata->init(dbs_data);
218 pr_err("%s: POLICY_INIT: init() failed\n", __func__);
223 if (!have_governor_per_policy())
224 WARN_ON(cpufreq_get_global_kobject());
226 rc = sysfs_create_group(get_governor_parent_kobj(policy),
227 get_sysfs_attr(dbs_data));
229 cdata->exit(dbs_data);
234 policy->governor_data = dbs_data;
236 /* policy latency is in nS. Convert it to uS first */
237 latency = policy->cpuinfo.transition_latency / 1000;
241 /* Bring kernel and HW constraints together */
242 dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate,
243 MIN_LATENCY_MULTIPLIER * latency);
244 set_sampling_rate(dbs_data, max(dbs_data->min_sampling_rate,
245 latency * LATENCY_MULTIPLIER));
247 if ((cdata->governor == GOV_CONSERVATIVE) &&
248 (!policy->governor->initialized)) {
249 struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
251 cpufreq_register_notifier(cs_ops->notifier_block,
252 CPUFREQ_TRANSITION_NOTIFIER);
255 if (!have_governor_per_policy())
256 cdata->gdbs_data = dbs_data;
259 case CPUFREQ_GOV_POLICY_EXIT:
260 if (!--dbs_data->usage_count) {
261 sysfs_remove_group(get_governor_parent_kobj(policy),
262 get_sysfs_attr(dbs_data));
264 if (!have_governor_per_policy())
265 cpufreq_put_global_kobject();
267 if ((dbs_data->cdata->governor == GOV_CONSERVATIVE) &&
268 (policy->governor->initialized == 1)) {
269 struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
271 cpufreq_unregister_notifier(cs_ops->notifier_block,
272 CPUFREQ_TRANSITION_NOTIFIER);
275 cdata->exit(dbs_data);
277 cdata->gdbs_data = NULL;
280 policy->governor_data = NULL;
284 cpu_cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
286 if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
287 cs_tuners = dbs_data->tuners;
288 cs_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
289 sampling_rate = cs_tuners->sampling_rate;
290 ignore_nice = cs_tuners->ignore_nice_load;
292 od_tuners = dbs_data->tuners;
293 od_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
294 sampling_rate = od_tuners->sampling_rate;
295 ignore_nice = od_tuners->ignore_nice_load;
296 od_ops = dbs_data->cdata->gov_ops;
297 io_busy = od_tuners->io_is_busy;
301 case CPUFREQ_GOV_START:
305 mutex_lock(&dbs_data->mutex);
307 for_each_cpu(j, policy->cpus) {
308 struct cpu_dbs_common_info *j_cdbs =
309 dbs_data->cdata->get_cpu_cdbs(j);
312 j_cdbs->cur_policy = policy;
313 j_cdbs->prev_cpu_idle = get_cpu_idle_time(j,
314 &j_cdbs->prev_cpu_wall, io_busy);
316 j_cdbs->prev_cpu_nice =
317 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
319 mutex_init(&j_cdbs->timer_mutex);
320 INIT_DEFERRABLE_WORK(&j_cdbs->work,
321 dbs_data->cdata->gov_dbs_timer);
325 * conservative does not implement micro like ondemand
326 * governor, thus we are bound to jiffes/HZ
328 if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
329 cs_dbs_info->down_skip = 0;
330 cs_dbs_info->enable = 1;
331 cs_dbs_info->requested_freq = policy->cur;
333 od_dbs_info->rate_mult = 1;
334 od_dbs_info->sample_type = OD_NORMAL_SAMPLE;
335 od_ops->powersave_bias_init_cpu(cpu);
338 mutex_unlock(&dbs_data->mutex);
340 /* Initiate timer time stamp */
341 cpu_cdbs->time_stamp = ktime_get();
343 gov_queue_work(dbs_data, policy,
344 delay_for_sampling_rate(sampling_rate), true);
347 case CPUFREQ_GOV_STOP:
348 if (dbs_data->cdata->governor == GOV_CONSERVATIVE)
349 cs_dbs_info->enable = 0;
351 gov_cancel_work(dbs_data, policy);
353 mutex_lock(&dbs_data->mutex);
354 mutex_destroy(&cpu_cdbs->timer_mutex);
355 cpu_cdbs->cur_policy = NULL;
357 mutex_unlock(&dbs_data->mutex);
361 case CPUFREQ_GOV_LIMITS:
362 mutex_lock(&cpu_cdbs->timer_mutex);
363 if (policy->max < cpu_cdbs->cur_policy->cur)
364 __cpufreq_driver_target(cpu_cdbs->cur_policy,
365 policy->max, CPUFREQ_RELATION_H);
366 else if (policy->min > cpu_cdbs->cur_policy->cur)
367 __cpufreq_driver_target(cpu_cdbs->cur_policy,
368 policy->min, CPUFREQ_RELATION_L);
369 dbs_check_cpu(dbs_data, cpu);
370 mutex_unlock(&cpu_cdbs->timer_mutex);
375 EXPORT_SYMBOL_GPL(cpufreq_governor_dbs);