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 #include <linux/cpu.h>
30 
31 #include "cpufreq_governor.h"
32 
33 static struct kobject *get_governor_parent_kobj(struct cpufreq_policy *policy)
34 {
35 	if (have_governor_per_policy())
36 		return &policy->kobj;
37 	else
38 		return cpufreq_global_kobject;
39 }
40 
41 static struct attribute_group *get_sysfs_attr(struct dbs_data *dbs_data)
42 {
43 	if (have_governor_per_policy())
44 		return dbs_data->cdata->attr_group_gov_pol;
45 	else
46 		return dbs_data->cdata->attr_group_gov_sys;
47 }
48 
49 static inline u64 get_cpu_idle_time_jiffy(unsigned int cpu, u64 *wall)
50 {
51 	u64 idle_time;
52 	u64 cur_wall_time;
53 	u64 busy_time;
54 
55 	cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
56 
57 	busy_time = kcpustat_cpu(cpu).cpustat[CPUTIME_USER];
58 	busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_SYSTEM];
59 	busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_IRQ];
60 	busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_SOFTIRQ];
61 	busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_STEAL];
62 	busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_NICE];
63 
64 	idle_time = cur_wall_time - busy_time;
65 	if (wall)
66 		*wall = cputime_to_usecs(cur_wall_time);
67 
68 	return cputime_to_usecs(idle_time);
69 }
70 
71 u64 get_cpu_idle_time(unsigned int cpu, u64 *wall, int io_busy)
72 {
73 	u64 idle_time = get_cpu_idle_time_us(cpu, io_busy ? wall : NULL);
74 
75 	if (idle_time == -1ULL)
76 		return get_cpu_idle_time_jiffy(cpu, wall);
77 	else if (!io_busy)
78 		idle_time += get_cpu_iowait_time_us(cpu, wall);
79 
80 	return idle_time;
81 }
82 EXPORT_SYMBOL_GPL(get_cpu_idle_time);
83 
84 void dbs_check_cpu(struct dbs_data *dbs_data, int cpu)
85 {
86 	struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
87 	struct od_dbs_tuners *od_tuners = dbs_data->tuners;
88 	struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
89 	struct cpufreq_policy *policy;
90 	unsigned int max_load = 0;
91 	unsigned int ignore_nice;
92 	unsigned int j;
93 
94 	if (dbs_data->cdata->governor == GOV_ONDEMAND)
95 		ignore_nice = od_tuners->ignore_nice;
96 	else
97 		ignore_nice = cs_tuners->ignore_nice;
98 
99 	policy = cdbs->cur_policy;
100 
101 	/* Get Absolute Load (in terms of freq for ondemand gov) */
102 	for_each_cpu(j, policy->cpus) {
103 		struct cpu_dbs_common_info *j_cdbs;
104 		u64 cur_wall_time, cur_idle_time;
105 		unsigned int idle_time, wall_time;
106 		unsigned int load;
107 		int io_busy = 0;
108 
109 		j_cdbs = dbs_data->cdata->get_cpu_cdbs(j);
110 
111 		/*
112 		 * For the purpose of ondemand, waiting for disk IO is
113 		 * an indication that you're performance critical, and
114 		 * not that the system is actually idle. So do not add
115 		 * the iowait time to the cpu idle time.
116 		 */
117 		if (dbs_data->cdata->governor == GOV_ONDEMAND)
118 			io_busy = od_tuners->io_is_busy;
119 		cur_idle_time = get_cpu_idle_time(j, &cur_wall_time, io_busy);
120 
121 		wall_time = (unsigned int)
122 			(cur_wall_time - j_cdbs->prev_cpu_wall);
123 		j_cdbs->prev_cpu_wall = cur_wall_time;
124 
125 		idle_time = (unsigned int)
126 			(cur_idle_time - j_cdbs->prev_cpu_idle);
127 		j_cdbs->prev_cpu_idle = cur_idle_time;
128 
129 		if (ignore_nice) {
130 			u64 cur_nice;
131 			unsigned long cur_nice_jiffies;
132 
133 			cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE] -
134 					 cdbs->prev_cpu_nice;
135 			/*
136 			 * Assumption: nice time between sampling periods will
137 			 * be less than 2^32 jiffies for 32 bit sys
138 			 */
139 			cur_nice_jiffies = (unsigned long)
140 					cputime64_to_jiffies64(cur_nice);
141 
142 			cdbs->prev_cpu_nice =
143 				kcpustat_cpu(j).cpustat[CPUTIME_NICE];
144 			idle_time += jiffies_to_usecs(cur_nice_jiffies);
145 		}
146 
147 		if (unlikely(!wall_time || wall_time < idle_time))
148 			continue;
149 
150 		load = 100 * (wall_time - idle_time) / wall_time;
151 
152 		if (dbs_data->cdata->governor == GOV_ONDEMAND) {
153 			int freq_avg = __cpufreq_driver_getavg(policy, j);
154 			if (freq_avg <= 0)
155 				freq_avg = policy->cur;
156 
157 			load *= freq_avg;
158 		}
159 
160 		if (load > max_load)
161 			max_load = load;
162 	}
163 
164 	dbs_data->cdata->gov_check_cpu(cpu, max_load);
165 }
166 EXPORT_SYMBOL_GPL(dbs_check_cpu);
167 
168 static inline void __gov_queue_work(int cpu, struct dbs_data *dbs_data,
169 		unsigned int delay)
170 {
171 	struct cpu_dbs_common_info *cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
172 
173 	mod_delayed_work_on(cpu, system_wq, &cdbs->work, delay);
174 }
175 
176 void gov_queue_work(struct dbs_data *dbs_data, struct cpufreq_policy *policy,
177 		unsigned int delay, bool all_cpus)
178 {
179 	int i;
180 
181 	if (!all_cpus) {
182 		__gov_queue_work(smp_processor_id(), dbs_data, delay);
183 	} else {
184 		get_online_cpus();
185 		for_each_cpu(i, policy->cpus)
186 			__gov_queue_work(i, dbs_data, delay);
187 		put_online_cpus();
188 	}
189 }
190 EXPORT_SYMBOL_GPL(gov_queue_work);
191 
192 static inline void gov_cancel_work(struct dbs_data *dbs_data,
193 		struct cpufreq_policy *policy)
194 {
195 	struct cpu_dbs_common_info *cdbs;
196 	int i;
197 
198 	for_each_cpu(i, policy->cpus) {
199 		cdbs = dbs_data->cdata->get_cpu_cdbs(i);
200 		cancel_delayed_work_sync(&cdbs->work);
201 	}
202 }
203 
204 /* Will return if we need to evaluate cpu load again or not */
205 bool need_load_eval(struct cpu_dbs_common_info *cdbs,
206 		unsigned int sampling_rate)
207 {
208 	if (policy_is_shared(cdbs->cur_policy)) {
209 		ktime_t time_now = ktime_get();
210 		s64 delta_us = ktime_us_delta(time_now, cdbs->time_stamp);
211 
212 		/* Do nothing if we recently have sampled */
213 		if (delta_us < (s64)(sampling_rate / 2))
214 			return false;
215 		else
216 			cdbs->time_stamp = time_now;
217 	}
218 
219 	return true;
220 }
221 EXPORT_SYMBOL_GPL(need_load_eval);
222 
223 static void set_sampling_rate(struct dbs_data *dbs_data,
224 		unsigned int sampling_rate)
225 {
226 	if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
227 		struct cs_dbs_tuners *cs_tuners = dbs_data->tuners;
228 		cs_tuners->sampling_rate = sampling_rate;
229 	} else {
230 		struct od_dbs_tuners *od_tuners = dbs_data->tuners;
231 		od_tuners->sampling_rate = sampling_rate;
232 	}
233 }
234 
235 int cpufreq_governor_dbs(struct cpufreq_policy *policy,
236 		struct common_dbs_data *cdata, unsigned int event)
237 {
238 	struct dbs_data *dbs_data;
239 	struct od_cpu_dbs_info_s *od_dbs_info = NULL;
240 	struct cs_cpu_dbs_info_s *cs_dbs_info = NULL;
241 	struct od_ops *od_ops = NULL;
242 	struct od_dbs_tuners *od_tuners = NULL;
243 	struct cs_dbs_tuners *cs_tuners = NULL;
244 	struct cpu_dbs_common_info *cpu_cdbs;
245 	unsigned int sampling_rate, latency, ignore_nice, j, cpu = policy->cpu;
246 	int io_busy = 0;
247 	int rc;
248 
249 	if (have_governor_per_policy())
250 		dbs_data = policy->governor_data;
251 	else
252 		dbs_data = cdata->gdbs_data;
253 
254 	WARN_ON(!dbs_data && (event != CPUFREQ_GOV_POLICY_INIT));
255 
256 	switch (event) {
257 	case CPUFREQ_GOV_POLICY_INIT:
258 		if (have_governor_per_policy()) {
259 			WARN_ON(dbs_data);
260 		} else if (dbs_data) {
261 			dbs_data->usage_count++;
262 			policy->governor_data = dbs_data;
263 			return 0;
264 		}
265 
266 		dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
267 		if (!dbs_data) {
268 			pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
269 			return -ENOMEM;
270 		}
271 
272 		dbs_data->cdata = cdata;
273 		dbs_data->usage_count = 1;
274 		rc = cdata->init(dbs_data);
275 		if (rc) {
276 			pr_err("%s: POLICY_INIT: init() failed\n", __func__);
277 			kfree(dbs_data);
278 			return rc;
279 		}
280 
281 		rc = sysfs_create_group(get_governor_parent_kobj(policy),
282 				get_sysfs_attr(dbs_data));
283 		if (rc) {
284 			cdata->exit(dbs_data);
285 			kfree(dbs_data);
286 			return rc;
287 		}
288 
289 		policy->governor_data = dbs_data;
290 
291 		/* policy latency is in nS. Convert it to uS first */
292 		latency = policy->cpuinfo.transition_latency / 1000;
293 		if (latency == 0)
294 			latency = 1;
295 
296 		/* Bring kernel and HW constraints together */
297 		dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate,
298 				MIN_LATENCY_MULTIPLIER * latency);
299 		set_sampling_rate(dbs_data, max(dbs_data->min_sampling_rate,
300 					latency * LATENCY_MULTIPLIER));
301 
302 		if ((cdata->governor == GOV_CONSERVATIVE) &&
303 				(!policy->governor->initialized)) {
304 			struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
305 
306 			cpufreq_register_notifier(cs_ops->notifier_block,
307 					CPUFREQ_TRANSITION_NOTIFIER);
308 		}
309 
310 		if (!have_governor_per_policy())
311 			cdata->gdbs_data = dbs_data;
312 
313 		return 0;
314 	case CPUFREQ_GOV_POLICY_EXIT:
315 		if (!--dbs_data->usage_count) {
316 			sysfs_remove_group(get_governor_parent_kobj(policy),
317 					get_sysfs_attr(dbs_data));
318 
319 			if ((dbs_data->cdata->governor == GOV_CONSERVATIVE) &&
320 				(policy->governor->initialized == 1)) {
321 				struct cs_ops *cs_ops = dbs_data->cdata->gov_ops;
322 
323 				cpufreq_unregister_notifier(cs_ops->notifier_block,
324 						CPUFREQ_TRANSITION_NOTIFIER);
325 			}
326 
327 			cdata->exit(dbs_data);
328 			kfree(dbs_data);
329 			cdata->gdbs_data = NULL;
330 		}
331 
332 		policy->governor_data = NULL;
333 		return 0;
334 	}
335 
336 	cpu_cdbs = dbs_data->cdata->get_cpu_cdbs(cpu);
337 
338 	if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
339 		cs_tuners = dbs_data->tuners;
340 		cs_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
341 		sampling_rate = cs_tuners->sampling_rate;
342 		ignore_nice = cs_tuners->ignore_nice;
343 	} else {
344 		od_tuners = dbs_data->tuners;
345 		od_dbs_info = dbs_data->cdata->get_cpu_dbs_info_s(cpu);
346 		sampling_rate = od_tuners->sampling_rate;
347 		ignore_nice = od_tuners->ignore_nice;
348 		od_ops = dbs_data->cdata->gov_ops;
349 		io_busy = od_tuners->io_is_busy;
350 	}
351 
352 	switch (event) {
353 	case CPUFREQ_GOV_START:
354 		if (!policy->cur)
355 			return -EINVAL;
356 
357 		mutex_lock(&dbs_data->mutex);
358 
359 		for_each_cpu(j, policy->cpus) {
360 			struct cpu_dbs_common_info *j_cdbs =
361 				dbs_data->cdata->get_cpu_cdbs(j);
362 
363 			j_cdbs->cpu = j;
364 			j_cdbs->cur_policy = policy;
365 			j_cdbs->prev_cpu_idle = get_cpu_idle_time(j,
366 					       &j_cdbs->prev_cpu_wall, io_busy);
367 			if (ignore_nice)
368 				j_cdbs->prev_cpu_nice =
369 					kcpustat_cpu(j).cpustat[CPUTIME_NICE];
370 
371 			mutex_init(&j_cdbs->timer_mutex);
372 			INIT_DEFERRABLE_WORK(&j_cdbs->work,
373 					     dbs_data->cdata->gov_dbs_timer);
374 		}
375 
376 		/*
377 		 * conservative does not implement micro like ondemand
378 		 * governor, thus we are bound to jiffes/HZ
379 		 */
380 		if (dbs_data->cdata->governor == GOV_CONSERVATIVE) {
381 			cs_dbs_info->down_skip = 0;
382 			cs_dbs_info->enable = 1;
383 			cs_dbs_info->requested_freq = policy->cur;
384 		} else {
385 			od_dbs_info->rate_mult = 1;
386 			od_dbs_info->sample_type = OD_NORMAL_SAMPLE;
387 			od_ops->powersave_bias_init_cpu(cpu);
388 		}
389 
390 		mutex_unlock(&dbs_data->mutex);
391 
392 		/* Initiate timer time stamp */
393 		cpu_cdbs->time_stamp = ktime_get();
394 
395 		gov_queue_work(dbs_data, policy,
396 				delay_for_sampling_rate(sampling_rate), true);
397 		break;
398 
399 	case CPUFREQ_GOV_STOP:
400 		if (dbs_data->cdata->governor == GOV_CONSERVATIVE)
401 			cs_dbs_info->enable = 0;
402 
403 		gov_cancel_work(dbs_data, policy);
404 
405 		mutex_lock(&dbs_data->mutex);
406 		mutex_destroy(&cpu_cdbs->timer_mutex);
407 
408 		mutex_unlock(&dbs_data->mutex);
409 
410 		break;
411 
412 	case CPUFREQ_GOV_LIMITS:
413 		mutex_lock(&cpu_cdbs->timer_mutex);
414 		if (policy->max < cpu_cdbs->cur_policy->cur)
415 			__cpufreq_driver_target(cpu_cdbs->cur_policy,
416 					policy->max, CPUFREQ_RELATION_H);
417 		else if (policy->min > cpu_cdbs->cur_policy->cur)
418 			__cpufreq_driver_target(cpu_cdbs->cur_policy,
419 					policy->min, CPUFREQ_RELATION_L);
420 		dbs_check_cpu(dbs_data, cpu);
421 		mutex_unlock(&cpu_cdbs->timer_mutex);
422 		break;
423 	}
424 	return 0;
425 }
426 EXPORT_SYMBOL_GPL(cpufreq_governor_dbs);
427