1 /*
2  * drivers/cpufreq/cpufreq_governor.c
3  *
4  * CPUFREQ governors common code
5  *
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  *
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 
17 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
18 
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/tick.h>
27 #include <linux/types.h>
28 #include <linux/workqueue.h>
29 
30 #include "cpufreq_governor.h"
31 
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 
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)
65 		*wall = cputime_to_usecs(cur_wall_time);
66 
67 	return cputime_to_usecs(idle_time);
68 }
69 
70 u64 get_cpu_idle_time(unsigned int cpu, u64 *wall, int io_busy)
71 {
72 	u64 idle_time = get_cpu_idle_time_us(cpu, io_busy ? wall : NULL);
73 
74 	if (idle_time == -1ULL)
75 		return get_cpu_idle_time_jiffy(cpu, wall);
76 	else if (!io_busy)
77 		idle_time += get_cpu_iowait_time_us(cpu, wall);
78 
79 	return idle_time;
80 }
81 EXPORT_SYMBOL_GPL(get_cpu_idle_time);
82 
83 void dbs_check_cpu(struct dbs_data *dbs_data, int cpu)
84 {
85 	struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
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 
93 	if (dbs_data->cdata->governor == GOV_ONDEMAND)
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;
103 		u64 cur_wall_time, cur_idle_time;
104 		unsigned int idle_time, wall_time;
105 		unsigned int load;
106 		int io_busy = 0;
107 
108 		j_cdbs = dbs_data->cdata->get_cpu_cdbs(j);
109 
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);
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 
146 		if (unlikely(!wall_time || wall_time < idle_time))
147 			continue;
148 
149 		load = 100 * (wall_time - idle_time) / wall_time;
150 
151 		if (dbs_data->cdata->governor == GOV_ONDEMAND) {
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 
163 	dbs_data->cdata->gov_check_cpu(cpu, max_load);
164 }
165 EXPORT_SYMBOL_GPL(dbs_check_cpu);
166 
167 static inline void __gov_queue_work(int cpu, struct dbs_data *dbs_data,
168 		unsigned int delay)
169 {
170 	struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
171 
172 	mod_delayed_work_on(cpu, system_wq, &cdbs->work, delay);
173 }
174 
175 void gov_queue_work(struct dbs_data *dbs_data, struct cpufreq_policy *policy,
176 		unsigned int delay, bool all_cpus)
177 {
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;
194 
195 	for_each_cpu(i, policy->cpus) {
196 		cdbs = dbs_data->cdata->get_cpu_cdbs(i);
197 		cancel_delayed_work_sync(&cdbs->work);
198 	}
199 }
200 
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 
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)
234 {
235 	struct dbs_data *dbs_data;
236 	struct od_cpu_dbs_info_s *od_dbs_info = NULL;
237 	struct cs_cpu_dbs_info_s *cs_dbs_info = NULL;
238 	struct od_ops *od_ops = NULL;
239 	struct od_dbs_tuners *od_tuners = NULL;
240 	struct cs_dbs_tuners *cs_tuners = NULL;
241 	struct cpu_dbs_common_info *cpu_cdbs;
242 	unsigned int sampling_rate, latency, ignore_nice, j, cpu = policy->cpu;
243 	int io_busy = 0;
244 	int rc;
245 
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) {
258 			policy->governor_data = dbs_data;
259 			return 0;
260 		}
261 
262 		dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
263 		if (!dbs_data) {
264 			pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
265 			return -ENOMEM;
266 		}
267 
268 		dbs_data->cdata = cdata;
269 		rc = cdata->init(dbs_data);
270 		if (rc) {
271 			pr_err("%s: POLICY_INIT: init() failed\n", __func__);
272 			kfree(dbs_data);
273 			return rc;
274 		}
275 
276 		rc = sysfs_create_group(get_governor_parent_kobj(policy),
277 				get_sysfs_attr(dbs_data));
278 		if (rc) {
279 			cdata->exit(dbs_data);
280 			kfree(dbs_data);
281 			return rc;
282 		}
283 
284 		policy->governor_data = dbs_data;
285 
286 		/* policy latency is in nS. Convert it to uS first */
287 		latency = policy->cpuinfo.transition_latency / 1000;
288 		if (latency == 0)
289 			latency = 1;
290 
291 		/* Bring kernel and HW constraints together */
292 		dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate,
293 				MIN_LATENCY_MULTIPLIER * latency);
294 		set_sampling_rate(dbs_data, max(dbs_data->min_sampling_rate,
295 					latency * LATENCY_MULTIPLIER));
296 
297 		if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
298 			struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
299 
300 			cpufreq_register_notifier(cs_ops->notifier_block,
301 					CPUFREQ_TRANSITION_NOTIFIER);
302 		}
303 
304 		if (!have_governor_per_policy())
305 			cdata->gdbs_data = dbs_data;
306 
307 		return 0;
308 	case CPUFREQ_GOV_POLICY_EXIT:
309 		if ((policy->governor->initialized == 1) ||
310 				have_governor_per_policy()) {
311 			sysfs_remove_group(get_governor_parent_kobj(policy),
312 					get_sysfs_attr(dbs_data));
313 
314 			if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
315 				struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
316 
317 				cpufreq_unregister_notifier(cs_ops->notifier_block,
318 						CPUFREQ_TRANSITION_NOTIFIER);
319 			}
320 
321 			cdata->exit(dbs_data);
322 			kfree(dbs_data);
323 			cdata->gdbs_data = NULL;
324 		}
325 
326 		policy->governor_data = NULL;
327 		return 0;
328 	}
329 
330 	cpu_cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
331 
332 	if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
333 		cs_tuners = dbs_data->tuners;
334 		cs_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
335 		sampling_rate = cs_tuners->sampling_rate;
336 		ignore_nice = cs_tuners->ignore_nice;
337 	} else {
338 		od_tuners = dbs_data->tuners;
339 		od_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
340 		sampling_rate = od_tuners->sampling_rate;
341 		ignore_nice = od_tuners->ignore_nice;
342 		od_ops = dbs_data->cdata->gov_ops;
343 		io_busy = od_tuners->io_is_busy;
344 	}
345 
346 	switch (event) {
347 	case CPUFREQ_GOV_START:
348 		if (!policy->cur)
349 			return -EINVAL;
350 
351 		mutex_lock(&dbs_data->mutex);
352 
353 		for_each_cpu(j, policy->cpus) {
354 			struct cpu_dbs_common_info *j_cdbs =
355 				dbs_data->cdata->get_cpu_cdbs(j);
356 
357 			j_cdbs->cpu = j;
358 			j_cdbs->cur_policy = policy;
359 			j_cdbs->prev_cpu_idle = get_cpu_idle_time(j,
360 					       &j_cdbs->prev_cpu_wall, io_busy);
361 			if (ignore_nice)
362 				j_cdbs->prev_cpu_nice =
363 					kcpustat_cpu(j).cpustat[CPUTIME_NICE];
364 
365 			mutex_init(&j_cdbs->timer_mutex);
366 			INIT_DEFERRABLE_WORK(&j_cdbs->work,
367 					     dbs_data->cdata->gov_dbs_timer);
368 		}
369 
370 		/*
371 		 * conservative does not implement micro like ondemand
372 		 * governor, thus we are bound to jiffes/HZ
373 		 */
374 		if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
375 			cs_dbs_info->down_skip = 0;
376 			cs_dbs_info->enable = 1;
377 			cs_dbs_info->requested_freq = policy->cur;
378 		} else {
379 			od_dbs_info->rate_mult = 1;
380 			od_dbs_info->sample_type = OD_NORMAL_SAMPLE;
381 			od_ops->powersave_bias_init_cpu(cpu);
382 		}
383 
384 		mutex_unlock(&dbs_data->mutex);
385 
386 		/* Initiate timer time stamp */
387 		cpu_cdbs->time_stamp = ktime_get();
388 
389 		gov_queue_work(dbs_data, policy,
390 				delay_for_sampling_rate(sampling_rate), true);
391 		break;
392 
393 	case CPUFREQ_GOV_STOP:
394 		if (dbs_data->cdata->governor == GOV_CONSERVATIVE)
395 			cs_dbs_info->enable = 0;
396 
397 		gov_cancel_work(dbs_data, policy);
398 
399 		mutex_lock(&dbs_data->mutex);
400 		mutex_destroy(&cpu_cdbs->timer_mutex);
401 
402 		mutex_unlock(&dbs_data->mutex);
403 
404 		break;
405 
406 	case CPUFREQ_GOV_LIMITS:
407 		mutex_lock(&cpu_cdbs->timer_mutex);
408 		if (policy->max < cpu_cdbs->cur_policy->cur)
409 			__cpufreq_driver_target(cpu_cdbs->cur_policy,
410 					policy->max, CPUFREQ_RELATION_H);
411 		else if (policy->min > cpu_cdbs->cur_policy->cur)
412 			__cpufreq_driver_target(cpu_cdbs->cur_policy,
413 					policy->min, CPUFREQ_RELATION_L);
414 		dbs_check_cpu(dbs_data, cpu);
415 		mutex_unlock(&cpu_cdbs->timer_mutex);
416 		break;
417 	}
418 	return 0;
419 }
420 EXPORT_SYMBOL_GPL(cpufreq_governor_dbs);
421