Merge branch 'cpufreq-fixes' of git://git.linaro.org/people/vireshk/linux into pm...
[linux-2.6-block.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 <linux/export.h>
20#include <linux/kernel_stat.h>
4d5dcc42 21#include <linux/slab.h>
4471a34f
VK
22
23#include "cpufreq_governor.h"
24
4d5dcc42
VK
25static struct attribute_group *get_sysfs_attr(struct dbs_data *dbs_data)
26{
27 if (have_governor_per_policy())
28 return dbs_data->cdata->attr_group_gov_pol;
29 else
30 return dbs_data->cdata->attr_group_gov_sys;
31}
32
4471a34f
VK
33void dbs_check_cpu(struct dbs_data *dbs_data, int cpu)
34{
4d5dcc42 35 struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
4471a34f
VK
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;
41 unsigned int j;
42
4d5dcc42 43 if (dbs_data->cdata->governor == GOV_ONDEMAND)
6c4640c3 44 ignore_nice = od_tuners->ignore_nice_load;
4471a34f 45 else
6c4640c3 46 ignore_nice = cs_tuners->ignore_nice_load;
4471a34f
VK
47
48 policy = cdbs->cur_policy;
49
dfa5bb62 50 /* Get Absolute Load */
4471a34f
VK
51 for_each_cpu(j, policy->cpus) {
52 struct cpu_dbs_common_info *j_cdbs;
9366d840
SK
53 u64 cur_wall_time, cur_idle_time;
54 unsigned int idle_time, wall_time;
4471a34f 55 unsigned int load;
9366d840 56 int io_busy = 0;
4471a34f 57
4d5dcc42 58 j_cdbs = dbs_data->cdata->get_cpu_cdbs(j);
4471a34f 59
9366d840
SK
60 /*
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.
65 */
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);
4471a34f
VK
69
70 wall_time = (unsigned int)
71 (cur_wall_time - j_cdbs->prev_cpu_wall);
72 j_cdbs->prev_cpu_wall = cur_wall_time;
73
74 idle_time = (unsigned int)
75 (cur_idle_time - j_cdbs->prev_cpu_idle);
76 j_cdbs->prev_cpu_idle = cur_idle_time;
77
78 if (ignore_nice) {
79 u64 cur_nice;
80 unsigned long cur_nice_jiffies;
81
82 cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] -
83 cdbs->prev_cpu_nice;
84 /*
85 * Assumption: nice time between sampling periods will
86 * be less than 2^32 jiffies for 32 bit sys
87 */
88 cur_nice_jiffies = (unsigned long)
89 cputime64_to_jiffies64(cur_nice);
90
91 cdbs->prev_cpu_nice =
92 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
93 idle_time += jiffies_to_usecs(cur_nice_jiffies);
94 }
95
4471a34f
VK
96 if (unlikely(!wall_time || wall_time < idle_time))
97 continue;
98
99 load = 100 * (wall_time - idle_time) / wall_time;
100
4471a34f
VK
101 if (load > max_load)
102 max_load = load;
103 }
104
4d5dcc42 105 dbs_data->cdata->gov_check_cpu(cpu, max_load);
4471a34f
VK
106}
107EXPORT_SYMBOL_GPL(dbs_check_cpu);
108
031299b3
VK
109static inline void __gov_queue_work(int cpu, struct dbs_data *dbs_data,
110 unsigned int delay)
4471a34f 111{
4d5dcc42 112 struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
4471a34f 113
031299b3 114 mod_delayed_work_on(cpu, system_wq, &cdbs->work, delay);
4471a34f
VK
115}
116
031299b3
VK
117void gov_queue_work(struct dbs_data *dbs_data, struct cpufreq_policy *policy,
118 unsigned int delay, bool all_cpus)
4471a34f 119{
031299b3
VK
120 int i;
121
122 if (!all_cpus) {
123 __gov_queue_work(smp_processor_id(), dbs_data, delay);
124 } else {
125 for_each_cpu(i, policy->cpus)
126 __gov_queue_work(i, dbs_data, delay);
127 }
128}
129EXPORT_SYMBOL_GPL(gov_queue_work);
130
131static inline void gov_cancel_work(struct dbs_data *dbs_data,
132 struct cpufreq_policy *policy)
133{
134 struct cpu_dbs_common_info *cdbs;
135 int i;
58ddcead 136
031299b3
VK
137 for_each_cpu(i, policy->cpus) {
138 cdbs = dbs_data->cdata->get_cpu_cdbs(i);
139 cancel_delayed_work_sync(&cdbs->work);
140 }
4471a34f
VK
141}
142
4447266b
VK
143/* Will return if we need to evaluate cpu load again or not */
144bool need_load_eval(struct cpu_dbs_common_info *cdbs,
145 unsigned int sampling_rate)
146{
147 if (policy_is_shared(cdbs->cur_policy)) {
148 ktime_t time_now = ktime_get();
149 s64 delta_us = ktime_us_delta(time_now, cdbs->time_stamp);
150
151 /* Do nothing if we recently have sampled */
152 if (delta_us < (s64)(sampling_rate / 2))
153 return false;
154 else
155 cdbs->time_stamp = time_now;
156 }
157
158 return true;
159}
160EXPORT_SYMBOL_GPL(need_load_eval);
161
4d5dcc42
VK
162static void set_sampling_rate(struct dbs_data *dbs_data,
163 unsigned int sampling_rate)
164{
165 if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
166 struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
167 cs_tuners->sampling_rate = sampling_rate;
168 } else {
169 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
170 od_tuners->sampling_rate = sampling_rate;
171 }
172}
173
174int cpufreq_governor_dbs(struct cpufreq_policy *policy,
175 struct common_dbs_data *cdata, unsigned int event)
4471a34f 176{
4d5dcc42 177 struct dbs_data *dbs_data;
4471a34f
VK
178 struct od_cpu_dbs_info_s *od_dbs_info = NULL;
179 struct cs_cpu_dbs_info_s *cs_dbs_info = NULL;
8eeed095 180 struct od_ops *od_ops = NULL;
4d5dcc42
VK
181 struct od_dbs_tuners *od_tuners = NULL;
182 struct cs_dbs_tuners *cs_tuners = NULL;
4471a34f 183 struct cpu_dbs_common_info *cpu_cdbs;
4d5dcc42 184 unsigned int sampling_rate, latency, ignore_nice, j, cpu = policy->cpu;
9366d840 185 int io_busy = 0;
4471a34f
VK
186 int rc;
187
4d5dcc42
VK
188 if (have_governor_per_policy())
189 dbs_data = policy->governor_data;
190 else
191 dbs_data = cdata->gdbs_data;
192
193 WARN_ON(!dbs_data && (event != CPUFREQ_GOV_POLICY_INIT));
194
195 switch (event) {
196 case CPUFREQ_GOV_POLICY_INIT:
197 if (have_governor_per_policy()) {
198 WARN_ON(dbs_data);
199 } else if (dbs_data) {
a97c98ad 200 dbs_data->usage_count++;
4d5dcc42
VK
201 policy->governor_data = dbs_data;
202 return 0;
203 }
204
205 dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
206 if (!dbs_data) {
207 pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
208 return -ENOMEM;
209 }
210
211 dbs_data->cdata = cdata;
a97c98ad 212 dbs_data->usage_count = 1;
4d5dcc42
VK
213 rc = cdata->init(dbs_data);
214 if (rc) {
215 pr_err("%s: POLICY_INIT: init() failed\n", __func__);
216 kfree(dbs_data);
217 return rc;
218 }
219
2361be23
VK
220 if (!have_governor_per_policy())
221 WARN_ON(cpufreq_get_global_kobject());
222
4d5dcc42
VK
223 rc = sysfs_create_group(get_governor_parent_kobj(policy),
224 get_sysfs_attr(dbs_data));
225 if (rc) {
226 cdata->exit(dbs_data);
227 kfree(dbs_data);
228 return rc;
229 }
230
231 policy->governor_data = dbs_data;
232
233 /* policy latency is in nS. Convert it to uS first */
234 latency = policy->cpuinfo.transition_latency / 1000;
235 if (latency == 0)
236 latency = 1;
237
238 /* Bring kernel and HW constraints together */
239 dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate,
240 MIN_LATENCY_MULTIPLIER * latency);
241 set_sampling_rate(dbs_data, max(dbs_data->min_sampling_rate,
242 latency * LATENCY_MULTIPLIER));
243
a97c98ad
VK
244 if ((cdata->governor == GOV_CONSERVATIVE) &&
245 (!policy->governor->initialized)) {
4d5dcc42
VK
246 struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
247
248 cpufreq_register_notifier(cs_ops->notifier_block,
249 CPUFREQ_TRANSITION_NOTIFIER);
250 }
251
252 if (!have_governor_per_policy())
253 cdata->gdbs_data = dbs_data;
254
255 return 0;
256 case CPUFREQ_GOV_POLICY_EXIT:
a97c98ad 257 if (!--dbs_data->usage_count) {
4d5dcc42
VK
258 sysfs_remove_group(get_governor_parent_kobj(policy),
259 get_sysfs_attr(dbs_data));
260
2361be23
VK
261 if (!have_governor_per_policy())
262 cpufreq_put_global_kobject();
263
a97c98ad
VK
264 if ((dbs_data->cdata->governor == GOV_CONSERVATIVE) &&
265 (policy->governor->initialized == 1)) {
4d5dcc42
VK
266 struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
267
268 cpufreq_unregister_notifier(cs_ops->notifier_block,
269 CPUFREQ_TRANSITION_NOTIFIER);
270 }
271
272 cdata->exit(dbs_data);
273 kfree(dbs_data);
274 cdata->gdbs_data = NULL;
275 }
4471a34f 276
4d5dcc42
VK
277 policy->governor_data = NULL;
278 return 0;
279 }
280
281 cpu_cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
282
283 if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
284 cs_tuners = dbs_data->tuners;
285 cs_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
286 sampling_rate = cs_tuners->sampling_rate;
6c4640c3 287 ignore_nice = cs_tuners->ignore_nice_load;
4471a34f 288 } else {
4d5dcc42
VK
289 od_tuners = dbs_data->tuners;
290 od_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
291 sampling_rate = od_tuners->sampling_rate;
6c4640c3 292 ignore_nice = od_tuners->ignore_nice_load;
4d5dcc42 293 od_ops = dbs_data->cdata->gov_ops;
9366d840 294 io_busy = od_tuners->io_is_busy;
4471a34f
VK
295 }
296
297 switch (event) {
298 case CPUFREQ_GOV_START:
3361b7b1 299 if (!policy->cur)
4471a34f
VK
300 return -EINVAL;
301
302 mutex_lock(&dbs_data->mutex);
303
4471a34f 304 for_each_cpu(j, policy->cpus) {
8eeed095 305 struct cpu_dbs_common_info *j_cdbs =
4d5dcc42 306 dbs_data->cdata->get_cpu_cdbs(j);
4471a34f 307
09dca5ae 308 j_cdbs->cpu = j;
4471a34f
VK
309 j_cdbs->cur_policy = policy;
310 j_cdbs->prev_cpu_idle = get_cpu_idle_time(j,
9366d840 311 &j_cdbs->prev_cpu_wall, io_busy);
4471a34f
VK
312 if (ignore_nice)
313 j_cdbs->prev_cpu_nice =
314 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
2abfa876
RA
315
316 mutex_init(&j_cdbs->timer_mutex);
317 INIT_DEFERRABLE_WORK(&j_cdbs->work,
4d5dcc42 318 dbs_data->cdata->gov_dbs_timer);
4471a34f
VK
319 }
320
4471a34f
VK
321 /*
322 * conservative does not implement micro like ondemand
323 * governor, thus we are bound to jiffes/HZ
324 */
4d5dcc42 325 if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
8eeed095
VK
326 cs_dbs_info->down_skip = 0;
327 cs_dbs_info->enable = 1;
328 cs_dbs_info->requested_freq = policy->cur;
4471a34f 329 } else {
8eeed095
VK
330 od_dbs_info->rate_mult = 1;
331 od_dbs_info->sample_type = OD_NORMAL_SAMPLE;
332 od_ops->powersave_bias_init_cpu(cpu);
4471a34f
VK
333 }
334
4471a34f
VK
335 mutex_unlock(&dbs_data->mutex);
336
58ddcead
FB
337 /* Initiate timer time stamp */
338 cpu_cdbs->time_stamp = ktime_get();
da53d61e 339
031299b3
VK
340 gov_queue_work(dbs_data, policy,
341 delay_for_sampling_rate(sampling_rate), true);
4471a34f
VK
342 break;
343
344 case CPUFREQ_GOV_STOP:
4d5dcc42 345 if (dbs_data->cdata->governor == GOV_CONSERVATIVE)
4471a34f
VK
346 cs_dbs_info->enable = 0;
347
031299b3 348 gov_cancel_work(dbs_data, policy);
4471a34f
VK
349
350 mutex_lock(&dbs_data->mutex);
351 mutex_destroy(&cpu_cdbs->timer_mutex);
419e1721 352 cpu_cdbs->cur_policy = NULL;
8eeed095 353
4471a34f
VK
354 mutex_unlock(&dbs_data->mutex);
355
356 break;
357
358 case CPUFREQ_GOV_LIMITS:
359 mutex_lock(&cpu_cdbs->timer_mutex);
360 if (policy->max < cpu_cdbs->cur_policy->cur)
361 __cpufreq_driver_target(cpu_cdbs->cur_policy,
362 policy->max, CPUFREQ_RELATION_H);
363 else if (policy->min > cpu_cdbs->cur_policy->cur)
364 __cpufreq_driver_target(cpu_cdbs->cur_policy,
365 policy->min, CPUFREQ_RELATION_L);
366 dbs_check_cpu(dbs_data, cpu);
367 mutex_unlock(&cpu_cdbs->timer_mutex);
368 break;
369 }
370 return 0;
371}
372EXPORT_SYMBOL_GPL(cpufreq_governor_dbs);