cpufreq: Simplify cpufreq_add_dev()
[deliverable/linux.git] / drivers / cpufreq / cpufreq_governor.c
CommitLineData
2aacdfff 1/*
2 * drivers/cpufreq/cpufreq_governor.c
3 *
4 * CPUFREQ governors common code
5 *
4471a34f
VK
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>
11 *
2aacdfff 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.
15 */
16
4471a34f
VK
17#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
18
2aacdfff 19#include <asm/cputime.h>
4471a34f
VK
20#include <linux/cpufreq.h>
21#include <linux/cpumask.h>
2aacdfff 22#include <linux/export.h>
23#include <linux/kernel_stat.h>
4471a34f 24#include <linux/mutex.h>
2aacdfff 25#include <linux/tick.h>
26#include <linux/types.h>
4471a34f
VK
27#include <linux/workqueue.h>
28
29#include "cpufreq_governor.h"
30
2aacdfff 31static inline u64 get_cpu_idle_time_jiffy(unsigned int cpu, u64 *wall)
32{
33 u64 idle_time;
34 u64 cur_wall_time;
35 u64 busy_time;
36
37 cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
38
39 busy_time = kcpustat_cpu(cpu).cpustat[CPUTIME_USER];
40 busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_SYSTEM];
41 busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_IRQ];
42 busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_SOFTIRQ];
43 busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_STEAL];
44 busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_NICE];
45
46 idle_time = cur_wall_time - busy_time;
47 if (wall)
a0e5af3c 48 *wall = cputime_to_usecs(cur_wall_time);
2aacdfff 49
a0e5af3c 50 return cputime_to_usecs(idle_time);
2aacdfff 51}
52
1e7586a1 53u64 get_cpu_idle_time(unsigned int cpu, u64 *wall)
2aacdfff 54{
55 u64 idle_time = get_cpu_idle_time_us(cpu, NULL);
56
57 if (idle_time == -1ULL)
58 return get_cpu_idle_time_jiffy(cpu, wall);
59 else
60 idle_time += get_cpu_iowait_time_us(cpu, wall);
61
62 return idle_time;
63}
64EXPORT_SYMBOL_GPL(get_cpu_idle_time);
4471a34f
VK
65
66void dbs_check_cpu(struct dbs_data *dbs_data, int cpu)
67{
68 struct cpu_dbs_common_info *cdbs = dbs_data->get_cpu_cdbs(cpu);
69 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
70 struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
71 struct cpufreq_policy *policy;
72 unsigned int max_load = 0;
73 unsigned int ignore_nice;
74 unsigned int j;
75
76 if (dbs_data->governor == GOV_ONDEMAND)
77 ignore_nice = od_tuners->ignore_nice;
78 else
79 ignore_nice = cs_tuners->ignore_nice;
80
81 policy = cdbs->cur_policy;
82
83 /* Get Absolute Load (in terms of freq for ondemand gov) */
84 for_each_cpu(j, policy->cpus) {
85 struct cpu_dbs_common_info *j_cdbs;
1e7586a1 86 u64 cur_wall_time, cur_idle_time, cur_iowait_time;
4471a34f
VK
87 unsigned int idle_time, wall_time, iowait_time;
88 unsigned int load;
89
90 j_cdbs = dbs_data->get_cpu_cdbs(j);
91
92 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
93
94 wall_time = (unsigned int)
95 (cur_wall_time - j_cdbs->prev_cpu_wall);
96 j_cdbs->prev_cpu_wall = cur_wall_time;
97
98 idle_time = (unsigned int)
99 (cur_idle_time - j_cdbs->prev_cpu_idle);
100 j_cdbs->prev_cpu_idle = cur_idle_time;
101
102 if (ignore_nice) {
103 u64 cur_nice;
104 unsigned long cur_nice_jiffies;
105
106 cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] -
107 cdbs->prev_cpu_nice;
108 /*
109 * Assumption: nice time between sampling periods will
110 * be less than 2^32 jiffies for 32 bit sys
111 */
112 cur_nice_jiffies = (unsigned long)
113 cputime64_to_jiffies64(cur_nice);
114
115 cdbs->prev_cpu_nice =
116 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
117 idle_time += jiffies_to_usecs(cur_nice_jiffies);
118 }
119
120 if (dbs_data->governor == GOV_ONDEMAND) {
121 struct od_cpu_dbs_info_s *od_j_dbs_info =
122 dbs_data->get_cpu_dbs_info_s(cpu);
123
124 cur_iowait_time = get_cpu_iowait_time_us(j,
125 &cur_wall_time);
126 if (cur_iowait_time == -1ULL)
127 cur_iowait_time = 0;
128
129 iowait_time = (unsigned int) (cur_iowait_time -
130 od_j_dbs_info->prev_cpu_iowait);
131 od_j_dbs_info->prev_cpu_iowait = cur_iowait_time;
132
133 /*
134 * For the purpose of ondemand, waiting for disk IO is
135 * an indication that you're performance critical, and
136 * not that the system is actually idle. So subtract the
137 * iowait time from the cpu idle time.
138 */
139 if (od_tuners->io_is_busy && idle_time >= iowait_time)
140 idle_time -= iowait_time;
141 }
142
143 if (unlikely(!wall_time || wall_time < idle_time))
144 continue;
145
146 load = 100 * (wall_time - idle_time) / wall_time;
147
148 if (dbs_data->governor == GOV_ONDEMAND) {
149 int freq_avg = __cpufreq_driver_getavg(policy, j);
150 if (freq_avg <= 0)
151 freq_avg = policy->cur;
152
153 load *= freq_avg;
154 }
155
156 if (load > max_load)
157 max_load = load;
158 }
159
160 dbs_data->gov_check_cpu(cpu, max_load);
161}
162EXPORT_SYMBOL_GPL(dbs_check_cpu);
163
2abfa876
RA
164bool dbs_sw_coordinated_cpus(struct cpu_dbs_common_info *cdbs)
165{
166 struct cpufreq_policy *policy = cdbs->cur_policy;
167
168 return cpumask_weight(policy->cpus) > 1;
169}
170EXPORT_SYMBOL_GPL(dbs_sw_coordinated_cpus);
171
4471a34f 172static inline void dbs_timer_init(struct dbs_data *dbs_data,
2abfa876
RA
173 struct cpu_dbs_common_info *cdbs,
174 unsigned int sampling_rate,
175 int cpu)
4471a34f
VK
176{
177 int delay = delay_for_sampling_rate(sampling_rate);
2abfa876 178 struct cpu_dbs_common_info *cdbs_local = dbs_data->get_cpu_cdbs(cpu);
4471a34f 179
2abfa876 180 schedule_delayed_work_on(cpu, &cdbs_local->work, delay);
4471a34f
VK
181}
182
183static inline void dbs_timer_exit(struct cpu_dbs_common_info *cdbs)
184{
185 cancel_delayed_work_sync(&cdbs->work);
186}
187
188int cpufreq_governor_dbs(struct dbs_data *dbs_data,
189 struct cpufreq_policy *policy, unsigned int event)
190{
191 struct od_cpu_dbs_info_s *od_dbs_info = NULL;
192 struct cs_cpu_dbs_info_s *cs_dbs_info = NULL;
193 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
194 struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
195 struct cpu_dbs_common_info *cpu_cdbs;
196 unsigned int *sampling_rate, latency, ignore_nice, j, cpu = policy->cpu;
197 int rc;
198
199 cpu_cdbs = dbs_data->get_cpu_cdbs(cpu);
200
201 if (dbs_data->governor == GOV_CONSERVATIVE) {
202 cs_dbs_info = dbs_data->get_cpu_dbs_info_s(cpu);
203 sampling_rate = &cs_tuners->sampling_rate;
204 ignore_nice = cs_tuners->ignore_nice;
205 } else {
206 od_dbs_info = dbs_data->get_cpu_dbs_info_s(cpu);
207 sampling_rate = &od_tuners->sampling_rate;
208 ignore_nice = od_tuners->ignore_nice;
209 }
210
211 switch (event) {
212 case CPUFREQ_GOV_START:
213 if ((!cpu_online(cpu)) || (!policy->cur))
214 return -EINVAL;
215
216 mutex_lock(&dbs_data->mutex);
217
218 dbs_data->enable++;
219 cpu_cdbs->cpu = cpu;
220 for_each_cpu(j, policy->cpus) {
221 struct cpu_dbs_common_info *j_cdbs;
222 j_cdbs = dbs_data->get_cpu_cdbs(j);
223
224 j_cdbs->cur_policy = policy;
225 j_cdbs->prev_cpu_idle = get_cpu_idle_time(j,
226 &j_cdbs->prev_cpu_wall);
227 if (ignore_nice)
228 j_cdbs->prev_cpu_nice =
229 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
2abfa876
RA
230
231 mutex_init(&j_cdbs->timer_mutex);
232 INIT_DEFERRABLE_WORK(&j_cdbs->work,
233 dbs_data->gov_dbs_timer);
4471a34f
VK
234 }
235
236 /*
237 * Start the timerschedule work, when this governor is used for
238 * first time
239 */
240 if (dbs_data->enable != 1)
241 goto second_time;
242
243 rc = sysfs_create_group(cpufreq_global_kobject,
244 dbs_data->attr_group);
245 if (rc) {
246 mutex_unlock(&dbs_data->mutex);
247 return rc;
248 }
249
250 /* policy latency is in nS. Convert it to uS first */
251 latency = policy->cpuinfo.transition_latency / 1000;
252 if (latency == 0)
253 latency = 1;
254
255 /*
256 * conservative does not implement micro like ondemand
257 * governor, thus we are bound to jiffes/HZ
258 */
259 if (dbs_data->governor == GOV_CONSERVATIVE) {
260 struct cs_ops *ops = dbs_data->gov_ops;
261
262 cpufreq_register_notifier(ops->notifier_block,
263 CPUFREQ_TRANSITION_NOTIFIER);
264
265 dbs_data->min_sampling_rate = MIN_SAMPLING_RATE_RATIO *
266 jiffies_to_usecs(10);
267 } else {
268 struct od_ops *ops = dbs_data->gov_ops;
269
270 od_tuners->io_is_busy = ops->io_busy();
271 }
272
273 /* Bring kernel and HW constraints together */
274 dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate,
275 MIN_LATENCY_MULTIPLIER * latency);
276 *sampling_rate = max(dbs_data->min_sampling_rate, latency *
277 LATENCY_MULTIPLIER);
278
279second_time:
280 if (dbs_data->governor == GOV_CONSERVATIVE) {
281 cs_dbs_info->down_skip = 0;
282 cs_dbs_info->enable = 1;
283 cs_dbs_info->requested_freq = policy->cur;
284 } else {
285 struct od_ops *ops = dbs_data->gov_ops;
286 od_dbs_info->rate_mult = 1;
287 od_dbs_info->sample_type = OD_NORMAL_SAMPLE;
288 ops->powersave_bias_init_cpu(cpu);
289 }
290 mutex_unlock(&dbs_data->mutex);
291
2abfa876 292 if (dbs_sw_coordinated_cpus(cpu_cdbs)) {
da53d61e
FB
293 /* Initiate timer time stamp */
294 cpu_cdbs->time_stamp = ktime_get();
295
2abfa876
RA
296 for_each_cpu(j, policy->cpus) {
297 struct cpu_dbs_common_info *j_cdbs;
298
299 j_cdbs = dbs_data->get_cpu_cdbs(j);
300 dbs_timer_init(dbs_data, j_cdbs,
301 *sampling_rate, j);
302 }
303 } else {
304 dbs_timer_init(dbs_data, cpu_cdbs, *sampling_rate, cpu);
305 }
4471a34f
VK
306 break;
307
308 case CPUFREQ_GOV_STOP:
309 if (dbs_data->governor == GOV_CONSERVATIVE)
310 cs_dbs_info->enable = 0;
311
2abfa876
RA
312 if (dbs_sw_coordinated_cpus(cpu_cdbs)) {
313 for_each_cpu(j, policy->cpus) {
314 struct cpu_dbs_common_info *j_cdbs;
315
316 j_cdbs = dbs_data->get_cpu_cdbs(j);
317 dbs_timer_exit(j_cdbs);
318 }
319 } else {
320 dbs_timer_exit(cpu_cdbs);
321 }
4471a34f
VK
322
323 mutex_lock(&dbs_data->mutex);
324 mutex_destroy(&cpu_cdbs->timer_mutex);
325 dbs_data->enable--;
326 if (!dbs_data->enable) {
327 struct cs_ops *ops = dbs_data->gov_ops;
328
329 sysfs_remove_group(cpufreq_global_kobject,
330 dbs_data->attr_group);
331 if (dbs_data->governor == GOV_CONSERVATIVE)
332 cpufreq_unregister_notifier(ops->notifier_block,
333 CPUFREQ_TRANSITION_NOTIFIER);
334 }
335 mutex_unlock(&dbs_data->mutex);
336
337 break;
338
339 case CPUFREQ_GOV_LIMITS:
340 mutex_lock(&cpu_cdbs->timer_mutex);
341 if (policy->max < cpu_cdbs->cur_policy->cur)
342 __cpufreq_driver_target(cpu_cdbs->cur_policy,
343 policy->max, CPUFREQ_RELATION_H);
344 else if (policy->min > cpu_cdbs->cur_policy->cur)
345 __cpufreq_driver_target(cpu_cdbs->cur_policy,
346 policy->min, CPUFREQ_RELATION_L);
347 dbs_check_cpu(dbs_data, cpu);
348 mutex_unlock(&cpu_cdbs->timer_mutex);
349 break;
350 }
351 return 0;
352}
353EXPORT_SYMBOL_GPL(cpufreq_governor_dbs);
This page took 0.062004 seconds and 5 git commands to generate.