Commit | Line | Data |
---|---|---|
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> |
4d5dcc42 | 25 | #include <linux/slab.h> |
2aacdfff | 26 | #include <linux/tick.h> |
27 | #include <linux/types.h> | |
4471a34f VK |
28 | #include <linux/workqueue.h> |
29 | ||
30 | #include "cpufreq_governor.h" | |
31 | ||
4d5dcc42 VK |
32 | static struct kobject *get_governor_parent_kobj(struct cpufreq_policy *policy) |
33 | { | |
34 | if (have_governor_per_policy()) | |
35 | return &policy->kobj; | |
36 | else | |
37 | return cpufreq_global_kobject; | |
38 | } | |
39 | ||
40 | static struct attribute_group *get_sysfs_attr(struct dbs_data *dbs_data) | |
41 | { | |
42 | if (have_governor_per_policy()) | |
43 | return dbs_data->cdata->attr_group_gov_pol; | |
44 | else | |
45 | return dbs_data->cdata->attr_group_gov_sys; | |
46 | } | |
47 | ||
2aacdfff | 48 | static inline u64 get_cpu_idle_time_jiffy(unsigned int cpu, u64 *wall) |
49 | { | |
50 | u64 idle_time; | |
51 | u64 cur_wall_time; | |
52 | u64 busy_time; | |
53 | ||
54 | cur_wall_time = jiffies64_to_cputime64(get_jiffies_64()); | |
55 | ||
56 | busy_time = kcpustat_cpu(cpu).cpustat[CPUTIME_USER]; | |
57 | busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_SYSTEM]; | |
58 | busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_IRQ]; | |
59 | busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_SOFTIRQ]; | |
60 | busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_STEAL]; | |
61 | busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_NICE]; | |
62 | ||
63 | idle_time = cur_wall_time - busy_time; | |
64 | if (wall) | |
a0e5af3c | 65 | *wall = cputime_to_usecs(cur_wall_time); |
2aacdfff | 66 | |
a0e5af3c | 67 | return cputime_to_usecs(idle_time); |
2aacdfff | 68 | } |
69 | ||
9366d840 | 70 | u64 get_cpu_idle_time(unsigned int cpu, u64 *wall, int io_busy) |
2aacdfff | 71 | { |
9366d840 | 72 | u64 idle_time = get_cpu_idle_time_us(cpu, io_busy ? wall : NULL); |
2aacdfff | 73 | |
74 | if (idle_time == -1ULL) | |
75 | return get_cpu_idle_time_jiffy(cpu, wall); | |
9366d840 | 76 | else if (!io_busy) |
2aacdfff | 77 | idle_time += get_cpu_iowait_time_us(cpu, wall); |
78 | ||
79 | return idle_time; | |
80 | } | |
81 | EXPORT_SYMBOL_GPL(get_cpu_idle_time); | |
4471a34f VK |
82 | |
83 | void dbs_check_cpu(struct dbs_data *dbs_data, int cpu) | |
84 | { | |
4d5dcc42 | 85 | struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu); |
4471a34f VK |
86 | struct od_dbs_tuners *od_tuners = dbs_data->tuners; |
87 | struct cs_dbs_tuners *cs_tuners = dbs_data->tuners; | |
88 | struct cpufreq_policy *policy; | |
89 | unsigned int max_load = 0; | |
90 | unsigned int ignore_nice; | |
91 | unsigned int j; | |
92 | ||
4d5dcc42 | 93 | if (dbs_data->cdata->governor == GOV_ONDEMAND) |
4471a34f VK |
94 | ignore_nice = od_tuners->ignore_nice; |
95 | else | |
96 | ignore_nice = cs_tuners->ignore_nice; | |
97 | ||
98 | policy = cdbs->cur_policy; | |
99 | ||
100 | /* Get Absolute Load (in terms of freq for ondemand gov) */ | |
101 | for_each_cpu(j, policy->cpus) { | |
102 | struct cpu_dbs_common_info *j_cdbs; | |
9366d840 SK |
103 | u64 cur_wall_time, cur_idle_time; |
104 | unsigned int idle_time, wall_time; | |
4471a34f | 105 | unsigned int load; |
9366d840 | 106 | int io_busy = 0; |
4471a34f | 107 | |
4d5dcc42 | 108 | j_cdbs = dbs_data->cdata->get_cpu_cdbs(j); |
4471a34f | 109 | |
9366d840 SK |
110 | /* |
111 | * For the purpose of ondemand, waiting for disk IO is | |
112 | * an indication that you're performance critical, and | |
113 | * not that the system is actually idle. So do not add | |
114 | * the iowait time to the cpu idle time. | |
115 | */ | |
116 | if (dbs_data->cdata->governor == GOV_ONDEMAND) | |
117 | io_busy = od_tuners->io_is_busy; | |
118 | cur_idle_time = get_cpu_idle_time(j, &cur_wall_time, io_busy); | |
4471a34f VK |
119 | |
120 | wall_time = (unsigned int) | |
121 | (cur_wall_time - j_cdbs->prev_cpu_wall); | |
122 | j_cdbs->prev_cpu_wall = cur_wall_time; | |
123 | ||
124 | idle_time = (unsigned int) | |
125 | (cur_idle_time - j_cdbs->prev_cpu_idle); | |
126 | j_cdbs->prev_cpu_idle = cur_idle_time; | |
127 | ||
128 | if (ignore_nice) { | |
129 | u64 cur_nice; | |
130 | unsigned long cur_nice_jiffies; | |
131 | ||
132 | cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] - | |
133 | cdbs->prev_cpu_nice; | |
134 | /* | |
135 | * Assumption: nice time between sampling periods will | |
136 | * be less than 2^32 jiffies for 32 bit sys | |
137 | */ | |
138 | cur_nice_jiffies = (unsigned long) | |
139 | cputime64_to_jiffies64(cur_nice); | |
140 | ||
141 | cdbs->prev_cpu_nice = | |
142 | kcpustat_cpu(j).cpustat[CPUTIME_NICE]; | |
143 | idle_time += jiffies_to_usecs(cur_nice_jiffies); | |
144 | } | |
145 | ||
4471a34f VK |
146 | if (unlikely(!wall_time || wall_time < idle_time)) |
147 | continue; | |
148 | ||
149 | load = 100 * (wall_time - idle_time) / wall_time; | |
150 | ||
4d5dcc42 | 151 | if (dbs_data->cdata->governor == GOV_ONDEMAND) { |
4471a34f VK |
152 | int freq_avg = __cpufreq_driver_getavg(policy, j); |
153 | if (freq_avg <= 0) | |
154 | freq_avg = policy->cur; | |
155 | ||
156 | load *= freq_avg; | |
157 | } | |
158 | ||
159 | if (load > max_load) | |
160 | max_load = load; | |
161 | } | |
162 | ||
4d5dcc42 | 163 | dbs_data->cdata->gov_check_cpu(cpu, max_load); |
4471a34f VK |
164 | } |
165 | EXPORT_SYMBOL_GPL(dbs_check_cpu); | |
166 | ||
031299b3 VK |
167 | static inline void __gov_queue_work(int cpu, struct dbs_data *dbs_data, |
168 | unsigned int delay) | |
4471a34f | 169 | { |
4d5dcc42 | 170 | struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu); |
4471a34f | 171 | |
031299b3 | 172 | mod_delayed_work_on(cpu, system_wq, &cdbs->work, delay); |
4471a34f VK |
173 | } |
174 | ||
031299b3 VK |
175 | void gov_queue_work(struct dbs_data *dbs_data, struct cpufreq_policy *policy, |
176 | unsigned int delay, bool all_cpus) | |
4471a34f | 177 | { |
031299b3 VK |
178 | int i; |
179 | ||
180 | if (!all_cpus) { | |
181 | __gov_queue_work(smp_processor_id(), dbs_data, delay); | |
182 | } else { | |
183 | for_each_cpu(i, policy->cpus) | |
184 | __gov_queue_work(i, dbs_data, delay); | |
185 | } | |
186 | } | |
187 | EXPORT_SYMBOL_GPL(gov_queue_work); | |
188 | ||
189 | static inline void gov_cancel_work(struct dbs_data *dbs_data, | |
190 | struct cpufreq_policy *policy) | |
191 | { | |
192 | struct cpu_dbs_common_info *cdbs; | |
193 | int i; | |
58ddcead | 194 | |
031299b3 VK |
195 | for_each_cpu(i, policy->cpus) { |
196 | cdbs = dbs_data->cdata->get_cpu_cdbs(i); | |
197 | cancel_delayed_work_sync(&cdbs->work); | |
198 | } | |
4471a34f VK |
199 | } |
200 | ||
4447266b VK |
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) | |
204 | { | |
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); | |
208 | ||
209 | /* Do nothing if we recently have sampled */ | |
210 | if (delta_us < (s64)(sampling_rate / 2)) | |
211 | return false; | |
212 | else | |
213 | cdbs->time_stamp = time_now; | |
214 | } | |
215 | ||
216 | return true; | |
217 | } | |
218 | EXPORT_SYMBOL_GPL(need_load_eval); | |
219 | ||
4d5dcc42 VK |
220 | static void set_sampling_rate(struct dbs_data *dbs_data, |
221 | unsigned int sampling_rate) | |
222 | { | |
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; | |
226 | } else { | |
227 | struct od_dbs_tuners *od_tuners = dbs_data->tuners; | |
228 | od_tuners->sampling_rate = sampling_rate; | |
229 | } | |
230 | } | |
231 | ||
232 | int cpufreq_governor_dbs(struct cpufreq_policy *policy, | |
233 | struct common_dbs_data *cdata, unsigned int event) | |
4471a34f | 234 | { |
4d5dcc42 | 235 | struct dbs_data *dbs_data; |
4471a34f VK |
236 | struct od_cpu_dbs_info_s *od_dbs_info = NULL; |
237 | struct cs_cpu_dbs_info_s *cs_dbs_info = NULL; | |
8eeed095 | 238 | struct od_ops *od_ops = NULL; |
4d5dcc42 VK |
239 | struct od_dbs_tuners *od_tuners = NULL; |
240 | struct cs_dbs_tuners *cs_tuners = NULL; | |
4471a34f | 241 | struct cpu_dbs_common_info *cpu_cdbs; |
4d5dcc42 | 242 | unsigned int sampling_rate, latency, ignore_nice, j, cpu = policy->cpu; |
9366d840 | 243 | int io_busy = 0; |
4471a34f VK |
244 | int rc; |
245 | ||
4d5dcc42 VK |
246 | if (have_governor_per_policy()) |
247 | dbs_data = policy->governor_data; | |
248 | else | |
249 | dbs_data = cdata->gdbs_data; | |
250 | ||
251 | WARN_ON(!dbs_data && (event != CPUFREQ_GOV_POLICY_INIT)); | |
252 | ||
253 | switch (event) { | |
254 | case CPUFREQ_GOV_POLICY_INIT: | |
255 | if (have_governor_per_policy()) { | |
256 | WARN_ON(dbs_data); | |
257 | } else if (dbs_data) { | |
a97c98ad | 258 | dbs_data->usage_count++; |
4d5dcc42 VK |
259 | policy->governor_data = dbs_data; |
260 | return 0; | |
261 | } | |
262 | ||
263 | dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL); | |
264 | if (!dbs_data) { | |
265 | pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__); | |
266 | return -ENOMEM; | |
267 | } | |
268 | ||
269 | dbs_data->cdata = cdata; | |
a97c98ad | 270 | dbs_data->usage_count = 1; |
4d5dcc42 VK |
271 | rc = cdata->init(dbs_data); |
272 | if (rc) { | |
273 | pr_err("%s: POLICY_INIT: init() failed\n", __func__); | |
274 | kfree(dbs_data); | |
275 | return rc; | |
276 | } | |
277 | ||
278 | rc = sysfs_create_group(get_governor_parent_kobj(policy), | |
279 | get_sysfs_attr(dbs_data)); | |
280 | if (rc) { | |
281 | cdata->exit(dbs_data); | |
282 | kfree(dbs_data); | |
283 | return rc; | |
284 | } | |
285 | ||
286 | policy->governor_data = dbs_data; | |
287 | ||
288 | /* policy latency is in nS. Convert it to uS first */ | |
289 | latency = policy->cpuinfo.transition_latency / 1000; | |
290 | if (latency == 0) | |
291 | latency = 1; | |
292 | ||
293 | /* Bring kernel and HW constraints together */ | |
294 | dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate, | |
295 | MIN_LATENCY_MULTIPLIER * latency); | |
296 | set_sampling_rate(dbs_data, max(dbs_data->min_sampling_rate, | |
297 | latency * LATENCY_MULTIPLIER)); | |
298 | ||
a97c98ad VK |
299 | if ((cdata->governor == GOV_CONSERVATIVE) && |
300 | (!policy->governor->initialized)) { | |
4d5dcc42 VK |
301 | struct cs_ops *cs_ops = dbs_data->cdata->gov_ops; |
302 | ||
303 | cpufreq_register_notifier(cs_ops->notifier_block, | |
304 | CPUFREQ_TRANSITION_NOTIFIER); | |
305 | } | |
306 | ||
307 | if (!have_governor_per_policy()) | |
308 | cdata->gdbs_data = dbs_data; | |
309 | ||
310 | return 0; | |
311 | case CPUFREQ_GOV_POLICY_EXIT: | |
a97c98ad | 312 | if (!--dbs_data->usage_count) { |
4d5dcc42 VK |
313 | sysfs_remove_group(get_governor_parent_kobj(policy), |
314 | get_sysfs_attr(dbs_data)); | |
315 | ||
a97c98ad VK |
316 | if ((dbs_data->cdata->governor == GOV_CONSERVATIVE) && |
317 | (policy->governor->initialized == 1)) { | |
4d5dcc42 VK |
318 | struct cs_ops *cs_ops = dbs_data->cdata->gov_ops; |
319 | ||
320 | cpufreq_unregister_notifier(cs_ops->notifier_block, | |
321 | CPUFREQ_TRANSITION_NOTIFIER); | |
322 | } | |
323 | ||
324 | cdata->exit(dbs_data); | |
325 | kfree(dbs_data); | |
326 | cdata->gdbs_data = NULL; | |
327 | } | |
4471a34f | 328 | |
4d5dcc42 VK |
329 | policy->governor_data = NULL; |
330 | return 0; | |
331 | } | |
332 | ||
333 | cpu_cdbs = dbs_data->cdata->get_cpu_cdbs(cpu); | |
334 | ||
335 | if (dbs_data->cdata->governor == GOV_CONSERVATIVE) { | |
336 | cs_tuners = dbs_data->tuners; | |
337 | cs_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu); | |
338 | sampling_rate = cs_tuners->sampling_rate; | |
4471a34f VK |
339 | ignore_nice = cs_tuners->ignore_nice; |
340 | } else { | |
4d5dcc42 VK |
341 | od_tuners = dbs_data->tuners; |
342 | od_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu); | |
343 | sampling_rate = od_tuners->sampling_rate; | |
4471a34f | 344 | ignore_nice = od_tuners->ignore_nice; |
4d5dcc42 | 345 | od_ops = dbs_data->cdata->gov_ops; |
9366d840 | 346 | io_busy = od_tuners->io_is_busy; |
4471a34f VK |
347 | } |
348 | ||
349 | switch (event) { | |
350 | case CPUFREQ_GOV_START: | |
3361b7b1 | 351 | if (!policy->cur) |
4471a34f VK |
352 | return -EINVAL; |
353 | ||
354 | mutex_lock(&dbs_data->mutex); | |
355 | ||
4471a34f | 356 | for_each_cpu(j, policy->cpus) { |
8eeed095 | 357 | struct cpu_dbs_common_info *j_cdbs = |
4d5dcc42 | 358 | dbs_data->cdata->get_cpu_cdbs(j); |
4471a34f | 359 | |
09dca5ae | 360 | j_cdbs->cpu = j; |
4471a34f VK |
361 | j_cdbs->cur_policy = policy; |
362 | j_cdbs->prev_cpu_idle = get_cpu_idle_time(j, | |
9366d840 | 363 | &j_cdbs->prev_cpu_wall, io_busy); |
4471a34f VK |
364 | if (ignore_nice) |
365 | j_cdbs->prev_cpu_nice = | |
366 | kcpustat_cpu(j).cpustat[CPUTIME_NICE]; | |
2abfa876 RA |
367 | |
368 | mutex_init(&j_cdbs->timer_mutex); | |
369 | INIT_DEFERRABLE_WORK(&j_cdbs->work, | |
4d5dcc42 | 370 | dbs_data->cdata->gov_dbs_timer); |
4471a34f VK |
371 | } |
372 | ||
4471a34f VK |
373 | /* |
374 | * conservative does not implement micro like ondemand | |
375 | * governor, thus we are bound to jiffes/HZ | |
376 | */ | |
4d5dcc42 | 377 | if (dbs_data->cdata->governor == GOV_CONSERVATIVE) { |
8eeed095 VK |
378 | cs_dbs_info->down_skip = 0; |
379 | cs_dbs_info->enable = 1; | |
380 | cs_dbs_info->requested_freq = policy->cur; | |
4471a34f | 381 | } else { |
8eeed095 VK |
382 | od_dbs_info->rate_mult = 1; |
383 | od_dbs_info->sample_type = OD_NORMAL_SAMPLE; | |
384 | od_ops->powersave_bias_init_cpu(cpu); | |
4471a34f VK |
385 | } |
386 | ||
4471a34f VK |
387 | mutex_unlock(&dbs_data->mutex); |
388 | ||
58ddcead FB |
389 | /* Initiate timer time stamp */ |
390 | cpu_cdbs->time_stamp = ktime_get(); | |
da53d61e | 391 | |
031299b3 VK |
392 | gov_queue_work(dbs_data, policy, |
393 | delay_for_sampling_rate(sampling_rate), true); | |
4471a34f VK |
394 | break; |
395 | ||
396 | case CPUFREQ_GOV_STOP: | |
4d5dcc42 | 397 | if (dbs_data->cdata->governor == GOV_CONSERVATIVE) |
4471a34f VK |
398 | cs_dbs_info->enable = 0; |
399 | ||
031299b3 | 400 | gov_cancel_work(dbs_data, policy); |
4471a34f VK |
401 | |
402 | mutex_lock(&dbs_data->mutex); | |
403 | mutex_destroy(&cpu_cdbs->timer_mutex); | |
8eeed095 | 404 | |
4471a34f VK |
405 | mutex_unlock(&dbs_data->mutex); |
406 | ||
407 | break; | |
408 | ||
409 | case CPUFREQ_GOV_LIMITS: | |
410 | mutex_lock(&cpu_cdbs->timer_mutex); | |
411 | if (policy->max < cpu_cdbs->cur_policy->cur) | |
412 | __cpufreq_driver_target(cpu_cdbs->cur_policy, | |
413 | policy->max, CPUFREQ_RELATION_H); | |
414 | else if (policy->min > cpu_cdbs->cur_policy->cur) | |
415 | __cpufreq_driver_target(cpu_cdbs->cur_policy, | |
416 | policy->min, CPUFREQ_RELATION_L); | |
417 | dbs_check_cpu(dbs_data, cpu); | |
418 | mutex_unlock(&cpu_cdbs->timer_mutex); | |
419 | break; | |
420 | } | |
421 | return 0; | |
422 | } | |
423 | EXPORT_SYMBOL_GPL(cpufreq_governor_dbs); |